Reusable GPU query batches and resident bounded k-NN/radius neighborhoods

[rsasaki0109]

Feature neighborhoods still require CPU search/results. Add reusable explicitly uploaded GPU query batches and owned resident nearest-k/radius rows on the existing GPU KD-tree, ready for subsequent device consumers.

k=0..=64 preserves CPU accumulator selection and stable ordered ties. Radius search preserves traversal order, counts all matches and exposes bounded-prefix overflow; complete CPU readback rejects incomplete rows. Count-only search, ownership/size/dispatch guards, explicit padded-transfer receipts and a downstream GPU consumer are covered. Target/query payloads stay resident between search passes; ordinary feature algorithms and ICP/GICP placement are unchanged.

Two isolated release surveys compare completed calls with all query uploads and padded readback/decoding included, plus separately reported resident completion with no readback. At 65,537 synthetic lattice queries, completed k=16 improves by 4.124x [3.615, 4.693] / 5.041x [4.044, 6.106], and radius search by 2.574x [2.262, 2.925] / 3.046x [2.889, 3.209]. Small batches and reversed-line radius cases favor CPU. Full tables and boundaries are in the note; no universal crossover or full feature speedup is claimed.

Validation: 467,364 exact CPU/brute k-NN selections, 199,846 radius prefix entries with exact counts, 524,416 concurrent selections, 131,074 large partial-dispatch self matches, live-output/zero-padding/metadata/ownership checks and a direct device consumer. All GPU features passed 196 tests, registration 35 unit tests plus six explicit GPU ICP/GICP tests, and the facade test passed. All-target/feature clippy with warnings denied, minimal/WASM builds, formatting, privacy and unchanged-path source audits passed. Browser execution is not claimed.

Builds on GPU correspondence foundations merged in PR #165; GPU FPFH/ISS/MLS computation remains a follow-up family.