WIP: CPU fallbacks for CUDA kernels

[RisingRedstone]

This PR introduces CPU implementations for coordinate and map building kernels to support execution on systems with limited GPU VRAM.

Context & Discussion: See issue #30 for the discussion on architectural alignment and if this is something worth adding.

Implementation Status

• hashmap_build_sparse_conv_out_coords (Implemented and tested)
• expand_unique_build_sparse_conv_out_coords (Implemented and tested)
• hashmap_build_sparse_conv_neighbour_map
• hashmap_build_submanifold_conv_neighbour_map

Current Optimizations & Performance Notes

• The initial CPU kernels prioritize correctness and roughly mirror the thread block semantics of the CUDA implementations.
• Masked Select Bypass: After some profiling, I found out that the torch implementation of masked_select was slowing the code down significantly. I implemented my own masked_select function which sped up the whole function by 2-2.5x.
• Future Roadmap: Pending approval of this integration approach, the next phase will involve deep profiling of the CPU path. I will refactor and flatten the loops in each cpu kernel implementation (to reduce integer divisions), but I am deferring those structural changes until the full pipeline can be profiled.

Notes for Reviewers

• This PR is currently in a Draft state to validate the integration approach before the final two map-building functions are written.