Add Intl.Segmenter modes: intl-word and grapheme

[krkarma777]

Closes #15

Summary

• `mode: 'intl-word'`: locale-aware word segmentation via `Intl.Segmenter` — unspaced scripts (Japanese/Chinese/Thai) now diff by word where `word` mode sees one opaque letter-run (私は猫が好きです → 私は犬が好きです isolates 猫→犬)
• `mode: 'grapheme'`: cluster-safe character diff — ZWJ emoji (👨‍👩‍👧) and combining sequences stay whole where `char` splits code points
• New `locale` option (BCP 47, defaults to runtime locale); Segmenter instances cached per granularity+locale; clear `TypeError` on runtimes without `Intl.Segmenter`
• `refine` drops `intl-word` pairs to grapheme granularity; `heuristic` composes
• Both modes route through the generic `diffTokens` pipeline; the fast scanner modes are untouched

Test Plan

• 7 new tests: Japanese word isolation (with the word-mode contrast), ZWJ replacement, refine-to-grapheme, 200 mixed-script fuzz round-trips, heuristic composition, empty inputs
• Full suite 50/50, typecheck clean (lib bumped to ES2022 for Segmenter types; emit target unchanged)