* move s2t data preprocess into paddlespeech.dataset

* avg model, compute wer, format rsl into paddlespeech.dataset

* fix format rsl

* fix avg ckpts