Dispatch preprocessing: port 6 more ops (bounds mask, coarse ijk, coords, inv index, avg pool, edge network)

[blackencino]

Part of #458 — stacked on #463

Review note: This PR is stacked on PR1 (#463). The diff includes PR1's changes since both target main.
After PR1 merges, this PR will be rebased and the diff will show only PR2 changes.

Summary

• Ported 6 additional ops to dispatch_table + new dispatch tools (14 dispatch tables total counting fwd+bwd):
  • ActiveVoxelsInBoundsMask — forEachActiveVoxel, device-only dispatch table
  • CoarseIjkForFineGrid — forEachActiveVoxel, device-only dispatch table (gains CPU support)
  • CoordsInGrid — dispatch::for_each + jagged_in (device x int_stype x contiguity)
  • IjkToInvIndex — dispatch::for_each + jagged_in (device x int_stype x contiguity)
  • DownsampleGridAvgPool (fwd+bwd) — forEachActiveVoxel over coarse grid with dual-grid accessor (device x float_stype x contiguity)
  • GridEdgeNetwork — forEachActiveVoxel, device-only, multi-output tensors
• Removed FVDB_DISPATCH_KERNEL_DEVICE from 9 call sites:
  • GridBatch.cpp (3): coordsInGrid, ijkToInvIndex, gridEdgeNetwork
  • GridBatchImpl.cu (1): activeVoxelsInBoundsMask
  • AvgPoolGrid.cpp (2): forward + backward
  • BuildCoarseGridFromFine.cu (2): CUDA + PrivateUse1 specializations
  • BuildGridForConv.cu (1): conv ijk shortcut path
• All op headers are now type-erased (no template <torch::DeviceType>)
• All precondition checks live inside the ops now, GridBatch calls to ops are thin.

New patterns established

• Device-only dispatch tables (1 axis) — ActiveVoxelsInBoundsMask, CoarseIjkForFineGrid, GridEdgeNetwork
• forEachJaggedElementChannel* proven unnecessary — CoordsInGrid and IjkToInvIndex use for_each + jagged_in directly (old code used forEachJaggedElementChannelCUDA/CPU with numChannels=1)
• Dual-grid voxel iteration — DownsampleGridAvgPool iterates coarse grid, reads fine grid via separate accessor (mirrors UpsampleGridNearest)
• Multi-output from forEachActiveVoxel — GridEdgeNetwork writes to 4 output tensors per voxel

Iteration pattern notes

• DownsampleGridAvgPool replaces per-(voxel, channel) GPU thread parallelism with a per-voxel sequential channel loop (matching UpsampleGridNearest). This trades per-channel parallelism for fewer redundant NanoVDB tree traversals in the pooling window. See comments in the op.
• ActiveVoxelsInBoundsMask loses a leaf-level hasOverlap(leaf.bbox()) early-out that the old code had. The old check was per-thread (not cooperative), and forEachActiveVoxel does not expose the leaf reference. Documented in comments; a future forEachLeaf-based variant could restore it.
• All other ops have identical GPU memory access patterns to the old code.

Test plan

• All gtests pass
• Python tests pass
• No old macros (FVDB_DISPATCH_KERNEL, AT_DISPATCH_V2) in any new/modified code