perf(features): parallelize ISS keypoints with bounded CPU workers

[rsasaki0109]

ISS detection previously evaluated every covariance and local maximum on the caller. feature-iss now shares one KD-tree across bounded scoped CPU workers, joining the saliency phase before selection reads it. Parallel selection writes a temporary one-byte-per-point mask and collects indices in input order; small/one-worker calls keep direct serial index collection. Public signatures, per-point math, equal-maxima behavior, schema, attributes and metadata are preserved.

Validation: 1,081,992 exact selection decisions and selected clouds against the frozen serial API; independent covariance eigenvalues, tied maxima and degenerate support; all six attribute buffer types including signed-zero/NaN words; configuration, empty/singleton and concurrent-call parity. Seven minimal and twenty-three all-feature native tests plus one facade smoke pass. Strict minimal/all-feature Clippy, crate/facade WASM compilation, formatting and diff checks pass.

Completed detect measurements include selected XYZ copying, using twelve conditions, two warmups and twenty rotating paired rounds. All six 32,768/65,536-point conditions favor public staging by 1.41–2.54x. Adverse one-worker controls at 16,384 and 32,768 noisy curved points are retained; serial neutrality and fixed cross-machine speedups are not claimed.

Execution boundary and all measurement conditions.