Optimize populated run-optimized BSI query results

[perfloop-agent]

Summary

The parallel BSI query paths (BatchEqual, CompareValue) called RunOptimize on every worker-local result bitmap while those bitmaps were still empty, paying per-worker optimization work that can never compress anything. This PR moves run optimization to the populated result:

• Worker-local result bitmaps are no longer run-optimized while empty; the aggregate returned by the parallel executor is optimized once, after all matches have been merged. BSI aggregate results get the same treatment.
• Empty results remain mutable, so callers can keep writing into them.
• The setter growth path no longer attempts to optimize an empty slice using an invalid index.

Populated contiguous results keep run compression exactly as before; only the wasted empty-bitmap passes are gone.

Performance

Measured with adjacent before/after builds, fresh benchmark processes, and isolated build caches:

workload	metric	before	after
CompareValue result serialization, 50k contiguous matches, 4 workers	ns/op	~2600	~187
same	B/op	9600	136
same	allocs/op	4	4

CompareValue end-to-end latency on the same workload was unchanged within noise, and BatchEqual moved ~30 ns/op on a ~1.7 ms operation — no material regression on any probed shape (alternating, sparse-container, single-match, and scattered variants).

Testing

• go test ./... and go test -race ./BitSliceIndexing pass.
• A differential probe covered signed values through the 64-bit boundaries, invalid and duplicate queries, the large scattered-query path, empty-result mutability, aggregate compression, and setter growth; all outputs and state checks match.

The branch includes a merge of current master; the only overlap was the shared BSI benchmark/test file, resolved by keeping the superset that contains both the existing BatchEqual coverage and the new run-optimized benchmarks.

Full verification record: https://app.perfloop.ai/t/oss/case_q4eynnf09a