feat: add deterministic MMR diversity reranker

[77cjj]

What & why

Adds a dependency-free MMRReranker and wires mmr / mmr:<lambda> into retrieval sweeps. It uses lexical relevance plus token-set Jaccard redundancy so users can measure relevance/diversity trade-offs without model downloads or network access.

Closes #7.

Checklist

• pytest -q passes (all tests; two existing optional-extra skips)
• ruff check src tests passes
• New unit tests cover near-duplicate diversification, stable deterministic ties, lambda validation, CLI parsing, sweep execution, and HTML reporting
• retrieval-lab demo runs green offline/keyless
• CHANGELOG.md updated
• README reranking options updated
• I read CONTRIBUTING.md

Notes

The implementation is deterministic and uses original shortlist order as the final tie-breaker, so no RNG state or seed is needed. ruff format --check src tests currently reports pre-existing formatting drift across 32 files on current main with Ruff 0.16.6; the files containing the new reranker and focused reranker tests were formatted, and ruff check is clean.

[AshwinUgale]

Thanks so much for this, @77cjj! 🙏Clean, well-scoped, and right in line with the project's
conventions — deterministic and keyless, dependency-free, with the full slice covered (CLI,
sweep, HTML, CHANGELOG, README) and thorough tests. Traced the MMR logic and it checks out.
Great addition — merging. 🎉