Dispatch preprocessing: new dispatch tools + 6 op ports + delete SimpleOpHelper

[blackencino]

Part of #458

Summary

• New src/fvdb/detail/dispatch/ module with dispatch-framework-aligned tools:
  • ForEachActiveVoxel.cuh — device-unified active-voxel iteration (built on dispatch::for_each)
  • GridAccessor.h — concept-constrained host/device accessor construction
  • JaggedView.h — jagged_in view for jagged tensors (mirrors dispatch::tensor_in)
  • TensorChecks.h — precondition validation utilities
• Ported 6 ops to dispatch_table + views (zero old macros):
  • ActiveGridGoords — forEachActiveVoxel + tensor_out
  • SerializeEncode — forEachActiveVoxel + tensor_out
  • MortonHilbertFromIjk — dispatch::for_each + tensor_in/tensor_out
  • IjkToIndex — dispatch::for_each + jagged_in + tensor_out (device x int stype x contiguity)
  • PointsInGrid — dispatch::for_each + jagged_in + tensor_out (device x float stype x contiguity)
• Deleted SimpleOpHelper.h (all 3 consumers ported, -332 lines)
• Removed FVDB_DISPATCH_KERNEL / FVDB_DISPATCH_KERNEL_DEVICE from 9 call sites in GridBatch.cpp and GridBatchImpl.cu
• All op headers are now type-erased (no template <torch::DeviceType>)

Test plan

• All 2218 unit tests pass (CPU + CUDA)
• No old macros (FVDB_DISPATCH_KERNEL, AT_DISPATCH_V2) in any new/modified code
• SimpleOpHelper.h has zero remaining includes