Add group capture support for Replace normalizer in Rust and Python

[ander-db]

Closes #1760

This PR adds support for backreferences (e.g., $1, $2, $0) in thereplacement string of the Replace normalizer, enabling pattern-basedreordering and transformation during normalization.

Background

PR #1788 looked abandoned, so I picked up the work. This implementation leverages the regex crate's native replacen/replace_all for $N expansion instead of re-running the regex engine on each match.

Key changes

• NormalizedString::replace_regex — new method (normalizer.rs:648) that accepts both a SysRegex (for finding matches — supports fancy-regex/onig features like lookahead) and an optional
  regex::Regex (for $1 expansion via replacen). Falls back to literal replacement when the pattern uses features regex::Regex can't compile (lookahead, etc.).
• Replace struct (replace.rs) — stores expansion_re: Option<regex::Regex> compiled from the raw regex
  pattern. normalize() passes it to replace_regex for alignment-tracked replacement. decode_chain() uses regex::Regex::replace_all when available, otherwise falls back to the literal find_matches loop.
• replace method — left completely unchanged (zero diff).
• onig.rs / fancy.rs — unmodified from original.

Backwards compatibility

• Patterns without capture groups or that use lookahead/lookbehind continue to work exactly as before (literal replacement).
• Serialization format is unchanged (expansion_re is #[serde(skip)]).

Tests

• Rust: 254 passed (202 unit + 21 doc + 31 integration), 0 failed
• Python: 198 passed, 3 skipped (pre-existing), 0 failures related to changes
  — test_replace_with_groups, test_replace_with_group_zero, test_replace_no_capture_unchanged all ✅

Note

My knowledge of Rust is limited, and this PR was developed with AI assistance. I've done my best to ensure correctness, but a careful review from someone more familiar with the codebase would be appreciated.