fix: F1 mask double-conversion NaN, F2 think token old format, S3 vocab sync, S1/S2 TODO, N1-N5

fix: v5 remove empty tokenizer.json, unify think tokens, integrate tokenizer into training pipeline