* fix mnist multi gpu save error; test=develop

* fix inference error; test=develop

* only run inference on rankid == 0; test=develop

test=develop

test=develop

* user new save load in all dygraph modell; test=develop

* change load_dict to set_dict; test=develop

* fix model save

* fix doc

* add ce for dygraph mnist

* add ce for dygraph mnist

* del mnist_dygraph.py

* change mnist_dygraph to train

* fix print style

* add multi process reader

* use distributed_batch_reader

* support multi-process for resnet and transformer

fix bug