Add GPU resident grid GICP covariance preparation

[rsasaki0109]

Depends on #174; CI must pass for the pinned source and dependency before merging.

Bounded resident-radius GICP adds source KD-tree preparation and fixed-capacity neighborhood buffers. Add explicit align_with_gpu_resident_grid_covariances / upload_with_gpu_grid_covariances for dense radius grids, gathering directly from original GPU positions into resident packed covariances. CPU counting-sort grids and the target tree remain, while the source tree, neighborhood rows and covariance readback/re-upload are eliminated.

Reuse the canonical grid arithmetic unchanged, with complete support capacity including self/duplicates. Overflow, malformed grid data or invalid SPD/numeric output zero both covariance sets and reject all equations. Grid indices use private scratch leased through one submission. Dense cell limits reject explicitly; existing sparse, k-NN and legacy routes retain their contracts. CPU 6x6 solves/pose updates remain.

Validation:

• GPU all-feature suite: 232 passing tests; registration all-feature library: 35 passing tests. Native grid alignment covers 162 scale/radius/stopping conditions, plus boundary/capacity/error/ownership/concurrency and minimal facade checks. Previous radius/k-NN regressions pass.
• 22,230 covariance rows / 133,380 f32 values match legacy grid output bit for bit; canonical shader and common pose/update loop are unchanged. Complete receipts and global invalidation are checked.
• GPU/registration all-target Clippy, CPU-only GICP, optional WASM compile, minimal facade and Rustfmt pass. No core/dependency/unsafe changes.
• Two isolated completed-call surveys, 20 cyclic samples per stage/condition. At 65,537 points, new grid is 1.18–1.47x faster than legacy GPU and 1.56–1.91x faster than bounded resident radius. All corresponding within-survey 95% intervals exceed one. Small host-update calls remain faster; no automatic placement/crossover claim is made.
• At that size, H2D is 3.96–4.42 MB (56.63–59.28% below legacy GPU); D2H stays 1,024 bytes. The duplicated capacity-128 condition no longer materializes its 135.3 MB neighborhood payload. Payload counts are not measured peak allocation.

Full conditions and every timing result: resident grid evidence.