perf(metrics): bound parallel Chamfer and Hausdorff queries

[rsasaki0109]

Chamfer/Hausdorff metrics previously queried and accumulated every point on the caller. metrics-distance now shares each target KD-tree across bounded scoped CPU workers, reuses a score window of at most 16,384 optional distances per worker and reuses each band's one-neighbor query buffer. Window joins precede original-order f64 accumulation, preserving exact metric words without regrouped worker sums. Small/one-worker directions retain the unchanged serial loop. Public signatures, metric definitions and empty/nonfinite behavior are preserved, with no dependency or device-transfer change.

Validation: 660 frozen f64 output words across staging thresholds, asymmetric clouds and forced 1/2/6 workers; independent asymmetric statistics/ties; a multi-window accumulation case that detects reassociation; NaN/None/infinity/empty behavior; renamed schema and concurrent calls. Eight native tests plus one facade smoke pass. Strict crate/facade Clippy, disabled-feature compilation, crate/facade WASM compilation, formatting and diff checks pass.

Completed two-direction API measurements include both KD-tree builds and use twelve conditions, two warmups and twenty rotating paired rounds. Final large-condition paired speedups are 1.14–1.85x, with all six intervals favoring public staging. The complete initial allocation-per-query survey and its adverse/uncertain controls are retained alongside the final survey. Small-input neutrality and fixed cross-machine speedups are not claimed. Optional-distance scratch is 128 KiB per worker on the measurement host; the two directions' windows are not live together.

Execution boundary and all measurements.