Add capture and replay support for VK_ARM_tensors and VK_ARM_data_graph

[fabian-lunarg]

Adds initial gfxreconstruct support for two ARM Vulkan extensions:

• VK_ARM_tensors: VkTensorARM / VkTensorViewARM objects and their memory
  binding, descriptor writes (VK_DESCRIPTOR_TYPE_TENSOR_ARM), and trimmed
  capture state reconstruction.
• VK_ARM_data_graph: VkDataGraphPipelineARM creation, VkDataGraphPipelineSessionARM
  creation/memory binding, and vkCmdDispatchDataGraphARM recording.

Both extensions are removed from kVulkanUnsupportedDeviceExtension so the
layer no longer strips them from device feature negotiation.

Capture (encode)

• Hand-written encoder for vkCreateDataGraphPipelinesARM to handle the
  FAIL_ON_COMPILE_REQUIRED early-out and the group-create pattern.
• PostProcess hooks for vkBindTensorMemoryARM,
  vkBindDataGraphPipelineSessionMemoryARM, and vkCreateTensorViewARM in
  VulkanCaptureManager.
• State tracker additions: TrackTensorMemoryBinding,
  TrackDataGraphPipelineSessionMemoryBinding,
  VK_DESCRIPTOR_TYPE_TENSOR_ARM descriptor tracking, DestroyState
  overloads, and VkObjectType debug name/tag handling for all three new
  types.
• kBindPoint_data_graph added to PipelineBindPoints.
• InitializeState / InitializeGroupObjectState specialisations for
  VkTensorARM, VkTensorViewARM, and VkDataGraphPipelineSessionARM.

Trimmed capture (state writer)

• WriteTensorState / WriteTensorMemoryState emit vkCreateTensorARM and
  vkBindTensorMemoryARM for all live tensors at trim start.
• WriteTensorSnapshotState / ProcessTensorMemory upload tensor contents
  via a new kInitTensorCommand metadata block
• WriteDataGraphPipelineSessionMemoryState reconstructs all session memory
  bindings, including recreation of pipelines (and their shader module /
  pipeline layout dependencies) that were destroyed before trim start.
• VK_DESCRIPTOR_TYPE_TENSOR_ARM descriptor write reconstruction in the
  descriptor set state writer.

Replay (decode)

• ProcessInitTensorCommand handler uploads tensor data via the resource
  initializer on trimmed-capture replay.
• OverrideCreateTensorARM, OverrideBindTensorMemoryARM,
  OverrideCreateDataGraphPipelineSessionARM,
  OverrideBindDataGraphPipelineSessionMemoryARM,
  OverrideDestroyTensorARM, OverrideDestroyDataGraphPipelineSessionARM,
  OverrideCreateDataGraphPipelinesARM, and
  OverrideGetTensorMemoryRequirementsARM added to the replay consumer.
• OverrideCreateDataGraphPipelinesARM checks optical-flow pipeline
  compatibility against the replay device's reported formats.
• Trimmed-capture replay adds TRANSFER_SRC/DST usage bits to tensor
  create infos so the resource initializer can upload snapshot data.

Allocator support

• VulkanDefaultAllocator: pass-through implementations for all new
  resource and memory-binding entry points.
• VulkanRebindAllocator: full rebind implementations for tensor and
  data graph session memory, including GetTensorMemoryUsage heuristic,
  AllocateMemoryForTensor (with VMA), and BindDataGraphPipelineSessionMemory
  (queries per-binding requirements from the driver at replay time).

Code generator

• Blacklists / replay_overrides updated for all hand-written entry points.
• Generator fixes for VkBaseOutStructure encoder/decoder wiring and
  decoder header template.
• Struct decoder, encoder, handle mapper, handle wrapper, deep-copy, and
  JSON generators extended for all new VK_ARM_tensors / VK_ARM_data_graph
  struct types.

Limitations

• Memory aliasing: Deferred to a follow-up PR.

fixes #2291

[MarkY-LunarG]

Additional comments from Claude:

  Code Review Findings

  [
    {
      "file": "framework/decode/vulkan_rebind_allocator.cpp",
      "line": 3875,
      "summary": "GetTensorMemoryRequirementsARM stores capture requirements from uninitialized struct before calling the Vulkan function that populates it.",
      "failure_scenario": "resource_alloc_info->capture_mem_reqs is assigned from memory_requirements->memoryRequirements at line 3875, but the Vulkan function 
  get_tensor_memory_requirements (which fills memory_requirements) is not called until line 3878. The stored capture requirements are whatever garbage was in the caller's 
  stack-allocated VkMemoryRequirements2. During replay, VMA uses these wrong requirements to allocate backing memory, causing allocation failures or size mismatches."
    },
    {
      "file": "framework/encode/vulkan_capture_manager.h",
      "line": 1779,
      "summary": "PostProcess_vkBindDataGraphPipelineSessionMemoryARM dereferences GetWrapper() result without a null check.",
      "failure_scenario": "GetWrapper() at line 1775 returns the DataGraphPipelineSessionARMWrapper for pBindInfos[i].session. If the session handle is invalid or not yet wrapped (e.g.
  trimmed capture resumes mid-frame), wrapper is null and wrapper->memory_bindings.begin() at line 1779 crashes. The analogous PostProcess_vkBindBufferMemory includes an equivalent 
  null guard."
    },
    {
      "file": "framework/decode/vulkan_rebind_allocator.cpp",
      "line": 3812,
      "summary": "BindTensorMemory silently masks per-tensor failures when binding multiple tensors: the result variable is overwritten by each iteration, so an early failure is erased
  if a later bind succeeds.",
      "failure_scenario": "With two tensors: AllocateMemoryForTensor fails for tensor[0] (result = VK_ERROR_*). The loop does not return early; iteration i=1 calls 
  AllocateMemoryForTensor and succeeds (result = VK_SUCCESS). The function returns VK_SUCCESS, and the caller assumes all binds succeeded. Contrast with 
  BindDataGraphPipelineSessionMemory, which returns immediately on error."
    },
    {
      "file": "framework/decode/vulkan_replay_consumer_base.cpp",
      "line": 13680,
      "summary": "Hint format compatibility check uses create_info.flowVectorFormat against info.hint_formats instead of a hint-specific format field.",
      "failure_scenario": "VkDataGraphPipelineOpticalFlowCreateInfoARM has no hintFormat field. hint_formats is populated by querying 
  VK_DATA_GRAPH_OPTICAL_FLOW_IMAGE_USAGE_HINT_BIT_ARM formats, which can differ from the output flow vector formats. The check at line 13680 validates whether the flow vector output 
  format is in the hint format list — semantically wrong. A pipeline where the hint image format differs from flowVectorFormat will either incorrectly fail compatibility or incorrectly
  pass, leading to a creation error or a wrong-device fallback at replay time."
    },
    {
      "file": "framework/encode/vulkan_state_tracker_initializers.h",
      "line": 743,
      "summary": "pDimensions is dereferenced inside a dimensionCount loop without a null guard.",
      "failure_scenario": "When TrackStruct processes a VkTensorDescriptionARM where dimensionCount > 0 but pDimensions == nullptr (malformed capture or fuzzing), the access 
  desc->pDimensions[i] at line 743 is a null pointer dereference. Per spec conventions dimensionCount > 0 implies non-null pDimensions, but this is a decode/state-restore path that 
  processes potentially untrusted .gfxr file data."
    },
    {
      "file": "framework/decode/vulkan_replay_consumer_base.cpp",
      "line": 13734,
      "summary": "data_graph_optical_flow_initialized flag is read and set without synchronization, creating a latent TOCTOU race.",
      "failure_scenario": "Two threads both hit OverrideCreateDataGraphPipelinesARM simultaneously on the same physical device for the first time. Both see 
  data_graph_optical_flow_initialized == false, both enter InitializeReplayDataGraphOpticalFlowInfo, both call clear() and then populate data_graph_optical_flow_infos concurrently, 
  resulting in a data race on the vector. Currently single-threaded replay prevents this, but it is a latent bug with no protection against future multi-threaded replay dispatch."
    }
  ]

Severity:

1. Critical
   2 - 3) High
   4-5) Medium
2. Low

Bugs 1–3 are the most actionable: #1 will silently corrupt tensor memory allocation sizes on every rebind-allocator replay, #2 is an unconditional crash if a handle is invalid, and #3 means multi-tensor bind errors go unreported.

[MarkY-LunarG]

We should create an issue to track this and mention it in a comment above the warning.

[fabian-lunarg]

I added the missing pieces (required additions to ResourceUtil, along with some consolidation-refactor).
so instead of an follow-up-issue there is now an additional commit which requires review

[fabian-lunarg]

Additional comments from Claude:

went through the list, applied corrections where required (some false positives)

[MarkY-LunarG]

Your new commit looks good too.