NanoVDB: own the small builders' device scratch with cuda::Buffer (CUDA)

[harrism]

Part of #2232 (step 2). Completes the raw-allocation audit that PointsToGrid (#2270), TopologyBuilder/MeshToGrid (#2268/#2269) and the topology-op consumers (#2273) already went through: the last ten mallocAsync/freeAsync pairs in the GPU tools become RAII cuda::Buffer scratch over an injectable resource.

What

• IndexToGrid (4 sites: the device NodeAccessor and the grid-name upload), addBlindData (2: byte-size scratch), GridStats (2: per-node statistics scratch), SignedFloodFill (2: node counts).
• Each gains a defaulted ResourceT template parameter following the established conventions: classes take a must-outlive resource instance in the constructor (as DilateGrid et al.), free functions take the template parameter and route through the per-type default instance (as pointsToGrid). Existing code is unaffected — every call site in the tree was traced and remains source-compatible.
• IndexToGrid's nested NodeAccessor is hoisted to a namespace-scope IndexToGridNodeAccessor<SrcBuildT> so the kernels that name it do not acquire a spurious ResourceT dependency; a public alias preserves the nested spelling.
• Behavior parity: every replaced free keeps its position in stream order (destroy(stream) at the old freeAsync site), and GridStats' mode-dependent allocation becomes a zero-element buffer when the mode has no average — the resource sees identical traffic, asserted exactly by the tests.

Tests

New seam tests in TestMemoryResource.cu with exact allocation counts per builder (IndexToGrid: 2, including the grid-name buffer; GridStats: 1, and 0 for the Extrema mode; SignedFloodFill: 1; addBlindData: 1 through the default-instance path), plus free-function ResourceT forwarding coverage.

Two observations for the reviewer

1. Pre-existing latent bug, preserved for parity: IndexToGrid::setGridName uploads the name to the device, but no kernel reads d_gridName — the name never reaches the output grid. This PR keeps that behavior (the upload now goes through the resource); fixing or removing it should be its own change.
2. Documented allocation boundary: GridStats::update's temporary NodeManager still allocates through the dual-space DeviceBuffer (createNodeManager is hasDeviceDual-gated), so that allocation bypasses an injected resource. Now called out in an @note on update(); the real fix belongs to the single-space handle work on NanoVDB: injectable CUDA memory resources — design & roadmap #2232.

Verification

• Zero warnings under the (now default) --Werror=all-warnings; full build of tests, tools and examples at CMAKE_CUDA_ARCHITECTURES=80.
• 7/7 ctests on a GPU runner (2× RTX 6000 Ada, CUDA 12.x).
• Reviewed by Codex (gpt-5.6-terra): one finding (the NodeManager boundary above), resolved; second pass clean with the reviewer independently building and running the test binary.

🤖 Generated with Claude Code

[kmuseth]

Seems like minor changes to adopt the new API for remaining tools. I am however confused as to why one member struct was changed to a namespaced struct

[kmuseth]

why is this no longer a member class of NodeAccesspr?

[harrism]

Nested types inherit all of the enclosing class's template parameters, so keeping it nested would make the kernels' (below) argument type vary with ResourceT and that would cause every kernel to be re-instantiated for every resource type. This would impact compile time and binary size.

Hoisting the struct outside the class keeps its type keyed on SrcBuildT only, with a member alias preserving the old spelling inside the class.

This is the same as TopologyBuilderData in #2268.