perf(vector/search): parallelize SegmentFanoutSearcher's per-segment fan-out (#926)

[mosuka]

Summary

Runs SegmentFanoutSearcher::search's per-segment probes concurrently on native targets (spun off from the #683 triage; rebased on #927/#928, whose shared-basis rescore it parallelizes along with each probe).

• The per-segment pipeline — concrete-searcher construction, the perf(vector/index): tiered merge policy + benches + docs for segment-per-commit (#634 PR-5) #883 expanding-refill loop, fix(vector/index): segmented search/merge dedup, sealed-segment deletes, PQ min-train policy (#634 PR-2) #880 newest-wins masking, the bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927 shared-basis rescore + min_similarity re-filter — is extracted into a self-contained probe_segment helper. Masking reads only immutable newer readers and the PreparedQuery is a shared read-only borrow, so one probe per worker thread is safe.
• Native: indexed into_par_iter().map(probe_segment).collect::<Result<_>>(), gated to readers.len() > 1 (single-segment indexes keep the zero-overhead serial call); wasm32 keeps the serial loop per the crate convention. The indexed collect preserves newest-first segment order, so the merged input to the bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927 distance sort is byte-identical to the serial path — results are bit-for-bit the same by construction.
• Rerank-basis segments (score_basis = "f32-rerank", bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927) skip the rescore inside their own probe task, exactly as in the serial code. Flat/IVF segmented indexes share the fan-out and get the parallelism for free.

Measured win

#683 probe fixture (10 segments, 20k vectors, dim 64, 100 queries, release, best of 3):

ef	serial (post-#927)	parallel	speedup
50	61.1ms	25.4ms	2.4x
100	88.6ms	38.1ms	2.3x
200	128.1ms	54.6ms	2.3x

Recall identical within build-RNG noise. This more than pays back #927's rescore overhead — 25.4ms is ~37% below even the pre-#927 serial baseline (40.1ms). The gap to core count comes from tiered-merge segment-size skew (the largest segment dominates the critical path).

Tests

• New segment_fanout_parallel_test.rs: repeated identical searches over a multi-segment fixture (stale duplicates + a deletion) return bit-identical results — parallel probing must not leak scheduling nondeterminism into the merged order.
• Cross-segment correctness under the parallel path is pinned by bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927's segment_score_comparability_test.rs (its brute-force fixture now exercises the concurrent fan-out on native) and the Stage-2 interaction by vector_rerank_engine_test.rs — both green unchanged.
• Single-segment indexes take the serial branch by construction, so no parallel-overhead regression is possible there.

Verification

• cargo fmt --check and cargo clippy --all-targets -- -D warnings clean on stable and 1.97.0
• cargo test -p laurus --lib: 1258 passed; --tests: all 64 binaries ok
• wasm32 cargo check -p laurus-wasm --target wasm32-unknown-unknown: clean (serial fallback)

Closes #926. Spun off from the #683 triage; depends on #927 (merged as #928).