remove the tmp buffer when merge grad

Implementing the MSRA initializer for rectifier units

* fix v2 init issue on Mac

* refine the v2 init

reorder parameters of layer

* fix

* Fix a bug

update resnet50 benchmark data

fix some warning with MKLDNN related codes and etc

Refine doc of pnpair evaluator

Fix prelu python api 

* Support many data types of several operators

* SeqConv only support float/double

* Revert adagrad

* Forward of raw if-else op

* add backward part of mnist if-else unittest

* refine fill_constant_batch_size_like layer

* add draft ifelse operator

* Complete IfElse Op

* add unittest of ifelse api

* merge baidu/develop

* Stash

* Merge develop branch

* Support int/int64 for fill_constant_batch_size_like

Update annotations of layers.py

refine MKLDNNLayer

remove ProtoData

* Improve the initializer Interface for fc, sequence_conv and conv2d layers
* Fix some typos in python code
* Fix CI

Add maxout operator.