Rule-based post accent correction via Yahoo MA POS metadata

[torrid-fish]

Problem

Regex-on-surface accent corrections cannot express grammar-based rules (for example "any verb in ます form gets a fall on ま") without enumerating every stem, because the pipeline carries no POS metadata. We need POS and conjugation info threaded into the word result, plus a rule/patch layer that can match on it, to fix suffix-level accents (動詞ます型, 形容詞たい型) and disambiguate homographs (升 as 名詞 vs ます as 助動詞).

[torrid-fish]

Resolved by PR #63. The POS-driven rule layer ships upstream-agnostic: apply_accent_patches (ます / たい patches), a pos_match predicate on the override engine, and POS columns (pos, pos1, conjugation_type, conjugation_form, base) on the word result, marked Field(exclude=True) so they stay out of the JSON. POS now comes from local UniDic (see #50), not Yahoo MA. Closes when #63 merges.