* when loss div batchsize, change lr, more epoch, loss can reduce more and cer lower than before

* since loss reduce more when loss div batchsize,  less lm alpha can be better.

* less lm alpha, more cer reduce

* alpha 2.2, cer 0.077478

* alpha 1.9, cer 0.077249

* large librispeech lr for batch_average ctc loss

* since loss reduce and model more confidence, then less lm alpha

* add acts, refactor ctc, add pos embed

* fix export, dataloader time log

* fix egs

* fix libri readme