* support fp16 training

* Use compiled training program

* Change timing ips.

* Use dali

* add pure fp16 training

* fix a bug, which will not use fuse pass using pure fp16 training.

* modify code as review

* modify loss, so that it will use different loss when using pure fp16 training.

* remove some fluid API

* add static optimizer.

* fix local rank
* fix export model

* add static running in dygraph

* unify model and model_name

* rm unsed archs fix bug of se_resnext

* fix distillation model for distributed.launch

* fix export model for googlenet

fix log benchmark

1. fix dist training 
2. fix cpp infer to support dir inference

polish codes and docs to adapt Paddle2.0-rc