Implement CUDA-style voxel grid pruning benchmark

[seanwevans]

Motivation

• Naive all-pairs close-approach search is O(N^2) and the roadmap specifies a GPU cell-grid pruning approach to reduce candidates.
• Provide a concrete, testable model of the roadmap’s pruning pipeline inside the benchmark backend to guide CUDA implementation and validate hashing/adjacency semantics.

Description

• Replaced the previous CUDA benchmark stub with simulate_cuda_cell_grid_conjunctions which models a uniform 3D voxel grid and per-object voxel hashing.
• Added GridConfig and Cartesian types plus coords_to_linear_hash and quantize_axis to assign stable voxel IDs to propagated positions.
• Implemented sorted voxel bucketing (analogue of a Thrust radix sort), compact voxel range construction, and candidate generation restricted to the same voxel and its 26 neighbors via are_neighbor_voxels and unflatten_hash.
• Wired the new path into the benchmark runner by calling the new simulation from CudaRunner and added focused unit tests for neighbor adjacency and hash round-trip stability.

Testing

• Ran cargo test -p openastrovizd and all unit tests passed (22 passed, 0 failed).
• Ran cargo fmt to format the changes and it completed successfully.

---

Codex Task