Add resident GPU GICP inverse weighting and normal-equation reductions

[rsasaki0109]

GICP's GPU correspondence route still downloaded every match and computed rotated covariance inverses/Hessian/gradient/fitness on CPU. Add align_with_gpu_reduction under register-gicp-gpu and reusable GpuGicpInput/owned normal equations under gpu-wgpu. Five device passes retain positions/covariances, transform originals, reuse unbounded nearest traversal and reduce inverse-weighted equations/error in one submission. Explicit readback is 256 bytes per evaluation; post-update equations serve the next solve/final fitness. CPU k-NN or explicit radius-GPU covariance readback/re-upload remains preparation, with complete receipts.

CPU solves centered/scaled 6x6 equations and converts the linearized step to the existing world update/stopping convention. Packed SPD validation, target-inverse trace normalization and relative covariance Cholesky guards reject unsupported numerics, invalidating all result payload without CPU fallback. Device f32 weights and normalized solving can change matching/pose/stopping; ordinary CPU and correspondence GICP source remains unchanged. Private scratch is leased through submission and owned results survive input drops/concurrent calls.

Two isolated release surveys retain 24 conditions and 20 paired samples per stage. Previous-GPU/new point estimates are 1.05-1.73x; two small-gain conditions have intervals including 1. Small inputs remain slower than the ordinary API. At 65,537 points, CPU k-NN preparation drops total D2H from 3,670,072 to 1,024 bytes; radius preparation still includes covariance readback, giving 4,195,392 total bytes versus 7,864,440 previously. Counts, raw fitness, precision differences, timing variability and bootstrap intervals are in the evidence note. Prepared-stage timings consistently use radius covariances and exclude preparation/readback/CPU solving.

Validation passed: 215 GPU tests, 35 registration library tests, two explicitly run adapter integrations covering 108 pose/scale/covariance/stopping conditions and 98,928 points, three existing GICP GPU tests and a facade test. Independent f64 covariance inversion/Jacobian/full H/g comparisons, world-basis conversion, strict conditioning/overflow/status, exact receipts, unequal clouds/outliers and concurrency pass. All-feature/CPU-only Clippy, minimal GPU, WASM and Rustfmt pass. Existing CPU/correspondence GICP and ICP device math is unchanged; current feature-isolation CI covers the new APIs.