fix: apply text replacements that overlap a match inside a word

[RaphaelFakhri]

Summary

ReplacementSet::replace collected candidates with Regex::find_iter and then dropped matches that fail the word-boundary check. find_iter never revisits a span it has already returned, so a rejected match inside a word hid any overlapping whole-phrase match that starts later in that span.

This affects phrases whose start repeats their end, such as stutter fixes:

• Rule I I -> I: Hi I I think so. stays unchanged, because i I inside Hi I is found first and rejected.
• Rule the the -> the: Breathe the the air. stays unchanged, because the the starting inside Breathe is found first and rejected.

The fix searches with Regex::find_at and, after a rejected match, resumes one character after its start. Accepted matches still advance past their end, so selection, ordering, and deletion behavior are unchanged.

Tests

• Adds a_match_inside_a_word_does_not_hide_an_overlapping_whole_phrase and a_rejected_match_ending_in_a_multibyte_character_resumes_on_a_character_boundary in src/text_replacements.rs.
• Both new tests fail without the fix (left: "Hi I I think so." and left: "Café é é.") and pass with it. All 7 text_replacements tests pass.
• cargo fmt --check and git diff --check are clean.

text_replacements is compiled only on macOS. The tests ran on Linux by compiling this exact module file in a scratch crate with a stub TextReplacement and regex = "=1.13.1", plus cargo clippy --all-targets -- -D warnings on that crate. The full macOS cargo test and Clippy run were not performed.

Fixes #139