* refine transformer cpu

* delete run_pretrain.sh

* support AMP training

* change the dir name

* add static step final save

* refine doc

* alter method name

* add benchmark readme

* add predict for static

* update doc for fleet api

* typo

* add dygraph amp support for transformer

* according params use_amp to decide use dygraph amp

* add fleet, test=develop

* support pad seq and bsz multi

* alter vocab size

* alter md5sum

* root for reader

* add static predict

* rm position_encoding_init

* fix static predict

* fix incorrect usage of data loader

* refine code

* remove unused code

* refine code

* add pad factor to pad vocab

* shuffle batch

* update reader

* update benchmark

* fix hung

* add shuffle batch

* update

* reader_seed to shuffle_seed

* seed for shuffle batch

* fix translation dataset and simplify transformer reader