fix: #448 drops inference mode during vLLM memory profiling

[shipiyouniao]

Problem

#448 replaces vLLM's GPUWorker.determine_available_memory() but does not
preserve the original method's @torch.inference_mode() decorator.

In vLLM 0.22.1, GPUModelRunner.profile_run() is not independently decorated
with @torch.inference_mode(); it relies on the outer worker method to establish
inference mode. As a result, the patched path runs the multimodal dummy profile
with autograd enabled, retains the vision encoder's backward graph and
intermediate activations, and can turn startup into a real CUDA OOM.

This is a regression introduced by #448 rather than an unavoidable multimodal
profiling peak.

Evidence

An A/B test used the same image, model, startup arguments and profiling inputs.
Only kvcached/integration/vllm/patches.py changed.

Both paths received the same inputs:

• V1 model runner
• pixel_values=(65536, 1536), bfloat16
• image_grid_thw=(1, 3)
• identical multimodal budgets and item counts

Memory observed around the same model.embed_multimodal() call:

Path	Before embed	After embed	PyTorch peak	Process footprint
Before #448	0.867 GiB	0.944 GiB	1.943 GiB	about 2.9 GiB
With #448	0.867 GiB	13.13 GiB	13.62 GiB	about 14.06 GiB

The increase occurs inside the vision encoder forward, before
determine_available_memory() returns and before KV cache configuration,
num_blocks, or _allocate_kv_cache_tensors() runs. This rules out eager KV
allocation, Mamba cache allocation, changed dummy inputs, and image resizing as
the cause.

The relevant execution contract is:

# vLLM 0.22.1
@torch.inference_mode()
def determine_available_memory(self) -> int:
    ...
    self.model_runner.profile_run()

The replacement introduced by #448 calls the same profiling methods without
that decorator. The expected mode difference is therefore:

• original path: torch.is_grad_enabled() == False and
  torch.is_inference_mode_enabled() == True
• patched path: torch.is_grad_enabled() == True and
  torch.is_inference_mode_enabled() == False

Proposed fix

Preserve vLLM's original execution contract by decorating the replacement with
@torch.inference_mode() (or entering the equivalent context) for the complete
profiling path.

This must not skip the model forward, multimodal profiling, CUDA graph
profiling/capture, or real CUDA error propagation. It only prevents autograd
state from being retained during an inference-only startup profile.

Regression tests should assert that both profile_run() and
profile_cudagraph_memory() execute with inference mode enabled when called
through the patched determine_available_memory().

Expected behavior

The patched worker should retain #448's logical KV-capacity behavior while
matching vLLM's original inference-mode semantics and avoiding the additional
vision-encoder activation peak.