feature/Add Broadcast and Gather op handle

Feature/mix cpu gpu op

Fix duplicate lr op after distribute transpiler

Speed/sequence expand

Fix Transformer Hang Problem

Add Wait() for reshape_op

Initialize large table value randomly

Refine double_buffer code

Multi stream thread pool

Split Executor.Run to Executor.Prepare and Executor.RunPreparedContext for inference

Fix warnings in sgd_op.h

Polish reshape op

fea/docker cudnn7

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

The fully connected: the operator is removed when the MKLDNN flag is OFF

* Fix CPPLint errors in operators

* Fix prior box op

* Fix Prior Box op

* Fix top_k_op.cu

* Fix pool mkmldnn

* Fix pool mkmldnn

update docker from cudnn5 -> cudnn7

Refine batch_norm_op.

Changing network configuration, avoid nan