NanoVDB: VoxelBlockManagerHandle accessors don't compile for buffer types whose data()/deviceData() return typed pointers

[swahtz]

Problem

nanovdb::tools::VoxelBlockManagerHandle<BufferT>'s eight pointer accessors static_cast directly from the buffer's data() / deviceData():

// nanovdb/tools/VoxelBlockManager.h:152 (and 159, 166, 173, 185, 188, 191, 194)
deviceFirstLeafID() { return static_cast<uint32_t*>(mFirstLeafID.deviceData()); }

This compiles when deviceData() returns void* (as NanoVDB's own HostBuffer / DeviceBuffer / UnifiedBuffer do), but a static_cast from any typed pointer to uint32_t* is ill-formed. A custom BufferT whose data() / deviceData() return e.g. uint8_t* fails to compile as soon as one of these accessors is instantiated:

nanovdb/tools/VoxelBlockManager.h(152): error: invalid type conversion
      deviceFirstLeafID() { return static_cast<uint32_t*>(mFirstLeafID.deviceData()); }
          detected during instantiation of ... VoxelBlockManagerHandle<BufferT>::deviceFirstLeafID()
          [with BufferT=fvdb::TorchDeviceBuffer, U=fvdb::TorchDeviceBuffer]

Nothing in the BufferTraits contract requires void* returns — and the rest of the generic surface (GridHandle, cuda::mergeGridHandles, buildVoxelBlockManager's BufferT::create path, etc.) works fine with typed-pointer buffers because a typed pointer converts implicitly to void*. VoxelBlockManagerHandle is the outlier: it's the only place that needs the conversion in the other direction and spells it as a single static_cast.

Where this bites

fvdb-core's TorchDeviceBuffer (its BufferT backed by PyTorch's caching allocator, uint8_t* data()/deviceData()) is used as the buffer type for grid handles throughout fvdb. Building a VoxelBlockManagerHandle<TorchDeviceBuffer> — the natural way to keep the VBM's firstLeafID / jumpMap arrays in the same memory pool as everything else, per the #2232 direction — fails to compile. fvdb currently works around it by allocating the arrays itself and launching BuildVoxelBlockManagerFunctor directly, bypassing the handle (openvdb/fvdb-core#733).

Proposed fix

Route the casts through void*, which is valid for both void* and typed-pointer buffers and is a no-op for the existing types:

deviceFirstLeafID() { return static_cast<uint32_t*>(static_cast<void*>(mFirstLeafID.deviceData())); }

(equivalently on the const overloads via const void*), applied to all eight accessors: hostFirstLeafID ×2, hostJumpMap ×2, deviceFirstLeafID ×2, deviceJumpMap ×2.

A TestNanoVDB-side guard could instantiate the handle over a minimal buffer whose data()/deviceData() return uint8_t* to keep this from regressing.

Happy to put up the one-file PR. Related context: the injectable-memory-resource roadmap #2232, where downstream buffer/resource types are expected to plug into these generic seams.