Add distribution implement of image classification

Add python wrapper for sequence reshape operator.

Add sequence reshape operator

* check if kernel if found for kernel type

* do kernel check before data transform

Change input data type to int64_t

* add data layout transform and optimize the implementation of data_transform

* add new_op_kernel_en.md

Add CTC align op

delete memory copy from linear_chain_crf_op.

Enhance GPU kernel of sequence erase op 

Add python doc for dynamic_lstm

* update new_op doc

* follow comment

Make merge and split support lodtensor

bring back test_parallel_do 

update doc and dist test due to transpiler change

Enhance distributed train performance

make auto-registry layers supporting specified output

* add split selected rows op

* update comment

* add grad check

* registry cuda kernel

* fix ci failed

Fix compiling warnings

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into distributed_split_selectedrows

update doc and dist test due to API change https://github.com/PaddlePaddle/Paddle/pull/7619#pullrequestreview-89656079

Remove optimize_op argument from get_pserver_program method of Distributed Transpiler

doc update with #7667

* update switch_kernel.md

* update the demo code

* update

Add feed and fetch op to ProgramDesc before saving for inference