[codex] Add AcuRank and BM25 top-20 benchmark

[pko89403]

Summary

• Add AcuRank as a built-in uncertainty-aware adaptive reranking strategy with sync/async support, TrueSkill-based scoring, score priors, strict validation, and an offline example.
• Add BM25 top-20 reranking benchmark support, method aliases, Pyserini candidate artifact handling, and committed benchmark summary evidence.
• Refresh README/README.ko, wiki/spec/reference docs, benchmark documentation, and bump the package version to 0.5.0.

Benchmark Evidence

• Dataset: AskUbuntuDupQuestions test data
• First-stage candidates: Pyserini BM25 top-20
• Evaluation: rerank top-20, evaluate @5, 361 queries
• Evidence artifact: benchmark-results/live/askubuntu-bm25-top20-default-live.v2.merged.json
• Candidate run artifact: benchmark-results/pyserini/askubuntu-bm25-top20.trec

Validation

• ./scripts/verify.sh -> 209 passed, lint/type/build succeeded
• UV_CACHE_DIR=.uv-cache uv tool run twine check dist/* -> passed for 0.5.0 wheel/sdist
• Pre-push hook reran verify + twine check successfully

Notes

• Raw *.query_results.jsonl benchmark rows and retry artifacts are intentionally ignored and not committed.
• Remaining invalid benchmark rows are reported in the summary artifact instead of being repaired.