MON-322 add reranking benchmark

[dommonkhouse]

Summary

• Add a reranking benchmark runner that compares plain hybrid MemSearch against optional ONNX reranking.
• Add scored fixture coverage for hit@1, hit@3, hit@5, duplicate-source regressions, per-query diffs, warm-up metadata, and conservative recommendations.
• Document benchmark-first rollout and rollback for reranker.model, including plain-mode isolation from global reranker config.

Test Plan

• env -u MEMSEARCH_DIR uv run python -m pytest -q
• uv run ruff check src tests scripts
• uv run ruff format --check src tests scripts
• uv run python scripts/benchmark_reranking.py --fixture-dir tests/fixtures/reranking --queries tests/fixtures/reranking/benchmark.json --collection ms_memsearch_ae2d4f9b --top-k 5 --reranker-model Alibaba-NLP/gte-reranker-modernbert-base --out /tmp/memsearch-reranking-fixture.json

Notes

• No global MemSearch config was changed.
• No deployment was performed.
• MEMSEARCH_DIR is cleared for pytest because the local shell sets it to a live shared state directory; leaving it set causes unrelated maintenance tests to share state.