fuse batch norm for conv operator 

- Added Epsilon (preventing softmax output from being too small) for softmax MKLDNN op

Fix conv_mkldnn_op.cc which is causing CI failure

* Fix cpplint warnings in beam_search_decode_op

* Fix cpplint warnings in beam_search_op

* Fix test

* fix

* Fix conv_op.h

* Fix conv_mkldnn_op

* Fix cpplint issues in detection_map_op

Add dist unitest data compare, ensure that dist train have same behavior as local training

Regenerate initializer.rst and add MSRAInitializer to initializer.__all__

- EPS added to softmax mkldnn primitive outcome is limited to training
phase

Fixes after review

clang format fixes

clang format fixes

Mpi enabled

Using constructor for VarHandle

move paddle/scripts/check_env.sh to benchmark/

The mkldnn tests are skipped when the MKLDNN flag is OFF

auto find tensorrt library and install in user root

Feature/tuning pe trans

Simplify DelayOps Logic

specify version for pip