…paths

- Add disk cleanup step to free ~25GB before running benchmarks
- Add msmarco_size input option (100K, 500K, 1M, full)
- Default to 1M passages for weekly runs (vs 8.8M full)
- Update download.sh to support subset creation

- Rewrite queries.sql to use EXPLAIN ANALYZE for proper index scan timing
- Add queries for different query types: short, medium, long, common, rare
- Add throughput benchmark (20 queries in batch)
- Add extract_metrics.sh to parse results into JSON format
- Add GitHub job summary with key metrics table
- Include benchmark_metrics.json in artifacts for historical comparison

- Fix grep patterns to extract index build and load times
- Fix RAISE NOTICE format specifiers for throughput output

The previous benchmark disabled index scans (enable_indexscan = off),
which meant it only tested sequential scan + scoring performance and
would not catch BM25 index regressions.

Changes:
- Remove enable_indexscan/enable_bitmapscan = off settings
- Add EXPLAIN ANALYZE for 5 representative queries to measure index
  scan latency (short, medium, long, common terms, rare terms)
- Add throughput test running all 225 standard Cranfield queries
- Add search quality validation (Precision@10 against relevance judgments)
- Add index statistics output

Local test results (1400 documents, 225 queries):
- Query latency: 3.5-6.6ms per query
- Throughput: 5.89 ms/query average (225 queries in 1.3s)
- Precision@10: 0.60

The gh-pages branch doesn't exist yet, so PR benchmarks and release
gates would fail trying to fetch it. Add skip-fetch-gh-pages: true
since these jobs don't save data anyway - they only compare/alert.

The gh-pages branch doesn't exist yet and the benchmark action fails
trying to switch to it. Switch to using GitHub Actions cache to store
benchmark history:

- PR benchmarks: restore from cache, compare, don't save (PRs don't
  update baseline)
- Full benchmarks (weekly/manual): restore from cache, compare, save
  new baseline to cache
- Release gate: restore from cache, compare with strict threshold

This approach:
1. Doesn't require gh-pages branch setup
2. Keeps benchmark history in Actions cache (persists across runs)
3. Still enables regression detection and PR comments