NanoVDB: type the Mask scratch buffers as Buffer<Mask<N>> once Mask is trivially copyable

[harrism]

Context

#2310 defaults nanovdb::Mask's copy constructor, making it trivially copyable. That satisfies cuda::Buffer<T, R>'s is_trivially_copyable<T> requirement, so the scratch buffers that exist solely to hold arrays of Mask<N> can become typed cuda::Buffer<Mask<N>, ...> instead of Buffer<std::byte> (or a legacy buffer) plus a cast. A census of the tree found exactly seven such sidecar sites; recording them here so they don't need re-finding.

Pure retypes (change the element type, delete the cast)

1. tools::cuda::MeshToGrid — the retainMaskBuffer in the first getHandle overload (allocated as bytes, reinterpret_cast to Mask<3>*, memset to 0xFF for all-on init).
2. tools::cuda::MeshToGrid — the per-subdivision-pass maskBuffer holding CTA hit results (Mask<3>). Its sibling countsBuffer is a separate allocation, so no offset arithmetic is affected.
3. tools::cuda::MeshToGrid — the retainMaskBuffer in the UDF-sidecar getHandle overload (same shape as 1).
4. examples/ex_dilate_nanovdb_cuda — dstLeafMaskBuffer in the kernels file (Mask<3> array).
5. unittest/TestNanoVDB.cu — the leaf-mask buffer in the dilation test (currently on the legacy dual-space buffer via create() + deviceData(); retyping folds into its migration to cuda::Buffer).

Retype with a small refactor

6. tools::cuda::TopologyBuilder::mUpperMasks (Mask<5> array): the builder's single byte-buffer ScratchT alias is shared by ~9 members, so typing the mask members means per-member aliases; the deviceUpperMasks() accessor returns void* and each consumer in util/cuda/Morphology.cuh and MeshToGrid.cuh casts — the accessor should return Mask<5>* and the casts go.
7. tools::cuda::TopologyBuilder::mLowerMasks (Mask<4>, viewed as a 2D Mask<4>(*)[Mask<5>::SIZE]): same alias/accessor refactor; note the 2D array-shape cast at the use site remains even after the element type is fixed.

Notes

• The zero/0xFF cudaMemsetAsync initializations stay as-is (a typed buffer's storage is still just bytes to initialize; noInit construction plus memset remains the right pattern).
• VoxelBlockManager has no mask sidecar despite being the intuitive candidate — its scratch arrays are uint32_t/uint64_t (typed-buffer candidates of their own, but not Mask).
• Sites 1–5 are all touched by NanoVDB: migrate off the dual-space buffers and deprecate the DeviceBuffer name (CUDA) #2301, so this work should land after it merges and be based on its post-merge shape to avoid conflicts.
• Out of scope: casts into grid internals (NanoVDB grids are byte blobs by design), shared-memory reinterpretations, and the union workaround in MeshToGrid's leaf-mask kernel — the latter is blocked by Mask's user-provided zeroing default constructor, which NanoVDB: default Mask's copy constructor #2310 deliberately leaves alone.