NanoVDB: type the remaining byte scratch buffers as cuda::Buffer<T> (CUDA)

[harrism]

Context

With #2312 / #2322 typing the Mask sidecars, a census of the CUDA tools found the same pattern — a cuda::Buffer<std::byte, R> allocated as N * sizeof(T) and reinterpreted as T* at every use — in ~25 more places. Typing them deletes the casts, replaces byte arithmetic with element counts (so Buffer's count×sizeof overflow check applies), and lets accessors return typed pointers. Recording the census so it doesn't need re-finding.

Already typed (verified): PointsToGrid, DistributedPointsToGrid, AddBlindData, IndexToGrid, GridStats, NodeManager's size scratch.

Pure retypes (homogeneous array for the buffer's whole life)

tools::cuda::TopologyBuilder (~30 casts)

• mUpperOffsets, mLowerOffsets, mLeafOffsets, mLowerParents, mLeafParents → Buffer<uint32_t>; mVoxelOffsets → Buffer<uint64_t> (five reinterpret_cast<uint64_t*> sites alone).
• The three local count buffers in the enumerate pass (upperCounts, lowerCounts, leafCounts) → Buffer<uint32_t>.
• mLowerOffsets, mLeafOffsets and the lower/leaf count locals are also read through a uint32_t (*)[Mask<5>::SIZE] view: give each a row-shaped accessor beside its allocation, the pattern deviceLowerMasks() established in NanoVDB: type the CUDA scratch buffers as cuda::Buffer<T> (CUDA) #2322, so the stride is fixed once.
• mDeviceData (one TopologyBuilderData) → Buffer<Data> of count 1, as DistributedPointsToGrid's mDataBuf already is.
• Once these are typed, the header's static_assert about "reinterpreting byte scratch as word-sized types" covers only the root staging and should be narrowed or removed.
• Stays bytes (correctly): mHostRoot/mDeviceRoot — a variable-sized root blob (RootT::memUsage(n) is not n * sizeof).

tools::cuda::MeshToGrid (~16 casts, no cross-file ripple)

• Members: mXformedTriangles (Triangle), mBoxTrianglePairsBuffer (MeshToGridBoxTrianglePair), mUniqueRootOriginsBuffer (Coord); their deviceXxx() accessors collapse to .data().
• Locals: rootBoxCounts, rootBoxOffsets, keysBuffer, sortedKeysBuffer, uniqueKeysBuffer, countsBuffer, offsetsBuffer (uint64_t), numSelectedBuffer (int32_t, count 1), newPairsBuffer (BoxTrianglePair).
• The byte ScratchT alias likely becomes unused in this file.

tools::cuda::SignedFloodFill: nodeBuffer (an array of RootChild<ValueT>) → Buffer<ChildT, ManagedResource>. The tree+root+tiles staging buffer beside it is heterogeneous and stays bytes.

API-level — needs a design decision, not a sweep

VoxelBlockManagerHandle<BufferT>: mFirstLeafID (uint32_t) and mJumpMap (uint64_t) drive ten accessor casts, but the handle takes one BufferT for both arrays. Typing them without changing the public signature means deriving the element-typed buffers from BufferT (e.g. via cuda::Buffer's rebind for the single-space family, with untyped fallback for HostBuffer/the dual family) rather than adding type parameters. Touches the host builder, the CUDA builder and TestNanoVDB; coordinate with the in-flight single-space VoxelBlockManagerHandle work. MeshToGrid's UDF sidecarBuffer is likewise a caller-supplied public return type (SidecarBufferT) and is out of scope for a sweep.

Correctly bytes — not candidates

GridHandle's chain-parse scratch (metadata array followed by a status struct), NodeManager's header-plus-three-arrays layout, TempPool (type-erased by construction), SignedFloodFill's tree/root/tiles staging, the tools' variable-sized root blobs (HostBuffer), and every cast into grid internals (grids are byte blobs by design).

[harrism]

The sweep portion (TopologyBuilder offsets/parents/counts/Data staging, all of MeshToGrid's scratch, SignedFloodFill's node buffer) is folded into #2322 alongside the Mask sidecars from #2312.

VoxelBlockManagerHandle stays open here, sequenced after #2291: that PR adds a cached leaf count to the handle and edits the same constructor, move operations and reset() lines a typed-buffer change would touch, and it is currently conflicting against master (its reset() edit collides with the destroy() dispatch from #2293). Landing the typing first would hand it a second conflict. The design when it does land: derive the two element-typed members from the existing BufferT via cuda::Buffer::rebind (with an untyped fallback for HostBuffer and the dual family), so the class template signature is unchanged — note the constructor's buffer parameter types do change, and the CUDA builder needs a typed allocation path since createDeviceStorage is byte-count based.

[harrism]

#2322 merged (0ab0a81): the sweep portion is done — TopologyBuilder's offsets/parents/counts/Data staging, all of MeshToGrid's scratch, SignedFloodFill's node buffer, plus the Mask sidecars from #2312. What remains here is only the VoxelBlockManagerHandle item, sequenced after #2291 as described above.