Keep GICP k-NN covariance preparation on GPU

[rsasaki0109]

Depends on #172 (resident point-to-plane reductions); this PR contains only the GICP k-NN covariance slice.

GICP's existing GPU reduction path still prepares k-NN covariances on CPU and uploads the source-sized covariance arrays. Add align_with_gpu_resident_covariances and GpuGicpInput::upload_with_gpu_covariances to compute bounded k-NN, centered/scaled sample covariance and plane eigenvectors on GPU, then retain those covariances for the existing equation/update loop. Original query/sample positions and target tree are reused; source-tree upload is counted. No neighborhood/covariance readback or covariance re-upload occurs.

CPU tree construction, normalized 6x6 solve and world pose updates remain. The new path supports k=3..=64, epsilon as f32 in 1e-4..=1 and covariance_radius=None. Existing CPU/uploaded/radius paths retain their behavior. Malformed neighborhoods or unsupported arithmetic invalidate all covariance and equation payload (equation status bit 16); f32 products/eigenvectors can change results at small/repeated spectra, without CPU fallback.

Validation:

• GPU library/integration: 224 passed; registration library: 35 passed. Five new native covariance tests check independent f64/spectral/Rayleigh math, malformed GPU rows, status propagation, limits, ownership and concurrent callers. Two new integrations compare 148,392 points in 162 pose/scale/k/stopping conditions, maximum scale-normalized pose-coordinate error 9.54e-7.
• Previous GPU reduction/radius integrations and facade smoke passed, as did all-feature/all-target and CPU-only Clippy, WASM compilation, minimal facade and formatting. Original CPU source and the shared evaluation/solve/pose loop are unchanged.
• Two isolated release surveys: 24 conditions, 20 samples per stage. At 65,537 points, completed calls improved 2.74–6.19x versus previous GPU and 4.29–6.64x versus CPU in this environment. Default k=20 gained 4.12x/4.36x versus previous GPU. H2D fell 37.9% (7,602,320 to 4,718,720 bytes); fixed D2H remains 1,024 bytes for three updates. Small 257-point calls favor CPU; one small previous/new interval includes one.

Full contracts, timing tables, transfers and precision limits.