Add GPU resident radius GICP covariance preparation

[rsasaki0109]

Radius covariance preparation currently downloads full covariance arrays and uploads them again for resident GICP equations. Add explicit align_with_gpu_resident_radius_covariances / upload_with_gpu_radius_covariances to keep complete inclusive radius neighborhoods, compensated covariance sums and plane eigenvectors on GPU. Reuse original query/sample positions and the target tree; retain the CPU tree construction and 6x6 solve/update boundary.

Caller capacity includes self/duplicates. Overflow or invalid arithmetic zeros both covariance sets and invalidates all equations; no truncated neighborhoods or automatic fallback are used. Sparse support retains epsilon*I. Existing CPU, grid and resident k-NN paths keep their contracts. The precision differences and measured latency tradeoffs are documented.

Validation:

• GPU all-feature suite: 227 passing tests; registration all-feature library: 35 passing tests. Native radius registration covers 162 scale/radius/stopping conditions, plus overflow, malformed rows, sparse extent beyond grid capacity, ownership/concurrency and minimal facade checks. Existing k-NN/reduction regressions pass.
• Independent brute-force/f64/spectral references: 13,320 rows; maximum unique-spectrum covariance-component error 1.51e-5. Maximum scale-normalized host-update pose-coordinate error is 9.54e-7.
• GPU/registration all-target Clippy, CPU-only GICP, optional WASM compile, minimal facade, benchmark Clippy and Rustfmt pass. No core/dependency/unsafe changes.
• Two isolated completed-call surveys, 20 cyclic samples per stage/condition. At 65,537 points, D2H falls from about 4.2 MB to 1,024 bytes and H2D falls 51.49–53.67%. Host-update/new latency ratios are 2.57–4.12x (host updates use GPU radius covariances), while previous-GPU/new ratios are 0.68–0.97x: no large-cloud latency improvement over the prior GPU implementation is claimed.

Full conditions, raw-sample locations, transfer tables and all timing results: resident radius evidence.