bench(native): add parallel sign-rq2 row matching OrdinalDB serving

[Fieldnote-Echo]

Summary

Add a native BEIR benchmark row for parallel sign-rq2 orchestration that mirrors the OrdinalDB serving path.

Why

PR #237's native benchmark harness currently gives HNSW a threaded path through hnsw.parallel_search, while the sign-rq2 row uses serial sign candidate generation plus serial subset rerank inside the batch loop. OrdinalDB's serving path now drives the ordvec 0.5 caller-owned APIs from rayon workers:

• split the query batch into chunks
• each worker calls SignBitmap::top_m_candidates_batched_serial_csr
• each worker reuses its own SubsetScratch
• each worker calls RankQuant::search_asymmetric_subset_batched_serial_into
• workers write directly into their output score/index chunks

Without this row, native benchmark comparisons answer "serial sign-rq2 vs threaded HNSW" rather than "ordvec two-stage vs HNSW under equivalent caller-owned threading."

Proposed benchmark row

Add a row such as ordvec-sign-rq2-par or sign-rq2-par to benchmarks/beir-bench:

1. Use the same corpus, query batch, k, candidate count, and thread pool size as the existing threaded comparison.
2. Build SignBitmap + RankQuant exactly like the current sign-rq2 row.
3. In the timed search closure, split query chunks across the configured rayon pool.
4. Per worker, allocate/reuse one SubsetScratch and call the serial CSR sign probe plus serial batched subset rerank into the assigned output slice.
5. Report the same recall/latency/build/storage columns as the existing rows.

Acceptance criteria

• The new row is apples-to-apples with HNSW for thread count and batch size.
• The existing serial sign-rq2 row remains, so single-thread/serial behavior is still visible.
• The benchmark summary makes clear which row is serial and which row is caller-parallel.
• The row uses public ordvec APIs only; no OrdinalDB dependency in the ordvec benchmark.

Related

• OrdinalDB 0.5 worktree now exposes DenseSearchPlan/DenseSearchTimings and uses caller-owned parallel sign→rerank orchestration.
• perf(api): add SignProbeScratch and allocation-reusing sign candidate APIs #234 tracks allocation-reusing sign probe APIs.
• perf(runtime): reuse scalar subset-rerank LUTs through SubsetScratch #235 tracks subset-rerank LUT reuse through SubsetScratch.
• perf(sign): explore tiled batched sign probe without full score matrix #236 tracks tiled batched sign probing without a full score matrix.
• diagnostics: optional per-call search execution reports #128 tracks native ordvec per-call execution reports.