Accelerate GPU voxel dispatch

[rsasaki0109]

What changed

• prefer a high-performance wgpu adapter by default while exposing an explicit low-power preference
• batch GPU bitonic-sort and prefix-scan stages into fewer command submissions
• replace large host zero-vector uploads with device-side buffer clears
• fix recycled voxel-key buffers so equal-sized later uploads retain COPY_DST compatibility
• keep centroid Auto on CPU until a new measured GPU crossover exists
• replace stale performance claims and charts with the 2026-07-16 receipt

Why

The previous runtime unconditionally requested a low-power adapter on hybrid systems. The end-to-end benchmark also exposed a 2M-point validation panic from incompatible pooled-buffer usage and substantial command-submission and host-allocation overhead. Meanwhile, later CPU fast paths invalidated the old README crossover claim.

Performance

Isolated Criterion GPU medians after the final batching change:

• 500k: 51.816 ms -> 35.829 ms (-30.9%)
• 1M: 78.258 ms -> 65.890 ms (-15.8%)
• 2M: 129.57 ms -> 104.82 ms (-19.1%)

The current CPU path remains faster through 2M, so Auto now stays on CPU for centroid mode. Explicit GPU and GPU-resident paths remain available. Public receipts intentionally omit the GPU model name.

Validation

• cargo test -p spatialrust-gpu --features gpu-wgpu (26 passed)
• cargo test -p spatialrust-filtering --features filter-voxel-gpu (18 passed)
• cargo clippy -p spatialrust-gpu --features gpu-wgpu --all-targets -- -D warnings
• cargo clippy -p spatialrust-filtering --features filter-voxel-gpu --all-targets -- -D warnings
• cargo fmt --all --check
• regenerated README benchmark and social SVGs