uniform row schema across template and llm rows (loadable with datasets)

card: valid task category

Multilingual TTS normalizer dataset: 52,698 written->spoken pairs, 16 locales + 6 Malaysian code-switched pairs

initial commit