fix: use converted token_ids for alignment for sensevoice model with timestamp output (#2429)

* fix: use converted token_ids for alignment

BPE doesn't guarantee converted ids (subwords) are revertible. which means `tokens` converted back is not always the same as `token_int`. A easy fix is to directly use the converted ids for alignment. Since they are from the same text, it shouldn't matter.

* fix: handle empty string

to index an empty string is to raise an exception. 这里没有判空。