Fix bug when memopt optimize loss.grad and use ParallelExecutor

Fix trt output name map

* Add Preface

* Add demo code

* Save file

* Refine code

* seems can work

* use elementwise strategy

* Use ElementwiseComputeEx

* Add comments

* extract functions from operator

* Refine code

* Follow comment

* code refine

* follow comments

* follow comments

add WITH_INFERENCE flag

dist lookup table only support sgd

* polish dist se resnext ut

* update

* update

* update

* avoid cpu initializer differ

* change to use executor for now

* update by comment

* remove lr decay use para exe, should fix para exe bug later

* update by comment

* fix the order of sum

* add doc

* check whether need to copy

* follow comments

Do not set loss@Grad as persistable

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into dist-lookup-table-only-support-sgd

avoid duplicated init op for one parameter

* fix the result of unit test

* fix resnext

* compare the result of PE and Exe

* compare the result of reduce and allreduce

* Simplify the travis script.
Now use docker to deploy documentations

* Check for the pull request

* Update paddle_build.sh

* Update paddle_build.sh

optimize profiler

Bugfix/disable anakin test

Revert part of a3ca4c99

Make lookup_table_op and softmax_op supporting high rank tensor

enforce test_parallel_executor_transformer and test_image_classification_resnet to run in serial (#12545)