feat: add LongMemEval benchmark script (was missing — published numbers weren't reproducible)

[anthroos]

Why

docs/benchmark-results.md (already on main) cites R@1 = 0.880, R@10 = 0.986, NDCG@10 = 0.924 and gives users a "Reproduce These Results" section with:

mkdir -p benchmarks/data
curl -L -o benchmarks/data/longmemeval_s_cleaned.json \
  "https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/..."

…but benchmarks/ doesn't exist in the repo. Anyone trying to reproduce ends with a downloaded dataset and no runner.

This commits the script behind the published numbers.

What

• benchmarks/longmemeval_bench.py — the runner. Three modes:
  • raw — pure vector similarity (baseline, like MemPalace raw)
  • hybrid — vector + BM25 keyword scoring (OpenExp default)
  • full — vector + BM25 + recency
  • Embedding: BAAI/bge-small-en-v1.5 (same as production)
  • In-memory Qdrant per question for clean isolation
  • Metrics computed the same way as MemPalace for apples-to-apples
• benchmarks/README.md — download URL, dataset format, three example invocations, expected runtime, expected numbers
• .gitignore — benchmarks/data/ added so the ~277 MB dataset can sit on disk without risk of accidental commit

No new dependencies. Uses fastembed and qdrant_client already in the project.

Test plan

• python benchmarks/longmemeval_bench.py benchmarks/data/longmemeval_s_cleaned.json --mode hybrid --limit 20 — should run end-to-end on 20 questions in a few minutes
• Full 500-question run in hybrid mode reproduces the numbers in docs/benchmark-results.md (R@1 ≈ 0.880, R@10 ≈ 0.986, NDCG@10 ≈ 0.924)