* update timit result, test=doc_fix

* result update

* fix bug

* add triplet loader

* empty preprocess file

* sync to u2, updating

* sync to u2 config

* fix bugs

* code refine

* update config

* customize decoding batch size

* update optimizer and lr scheduler

* minor

* minor

* minor

* fix bugs of refs

* minor

* distributed sampler

* minor

* refine the loader