Add match_addresses returning the winning address pair with alignment detail

[pudo]

Summary

• New rigour.addresses.match_addresses(queries, results) -> AddressMatch | None returns the best pair of a list×list address comparison with the evidence behind it: score, the raw query and result strings that won, and a one-line detail.
• detail is a diff-style token line: aligned tokens in query order (berlin, boulevard~blvd for fuzzy / keyword-alias / territory-code pairs), then -tok for query-only and +tok for result-only tokens. Example: sunset boulevard~blvd 12 -los -angeles +la.
• compare_address and compare_address_many share the same inner loop and keep returning only the score. Scores are unchanged.
• The alignment writes its bound pairs into a reused scratch buffer that is swapped into the winner's slot on improvement, so the hot N×M loop allocates nothing per pair and the detail string is built once, for the winner.

This lets nomenklatura's _address_match fill FtResult(score, detail, query, candidate) from one call instead of rebuilding its own detail text.

Test plan

• cargo test, cargo clippy --all-targets -- -D warnings (with and without --features python), cargo fmt --check
• pytest --cov rigour (535 passed), mypy --strict rigour
• mkdocs build --strict
• contrib/address_bench/evaluate.py --scorer rust: AUC 0.936, best threshold 0.292 (matches the documented ~0.3)

🤖 Generated with Claude Code