Use fixed activation in the lstm kernel, since there is some bug in the activation function pointer. It will be fixed later.

refine paddle_merge_model

* Cross Entropy Wrong

* Fix XE

* Polish gradient check for xe

* Fix compile

* Init

* Stash

* Polish SaveLoadOp

* Fix CI

* Polish code

* Save GPU Tensor

* Stash

* Fix CI

* Adding the L2 loss operator for L2 regularization
* Renaming l2_loss op to squared_l2_norm_op
* Addressing code review feedback

enable mkldnn_batch_norm

fix sparse update size

* Removing survey out of the design doc and fixing typos
* Fix Typos

"correct the description"

Fix pserver SGD with momemtum checkpoint recover.

* Adding nesterov momentum to python momentum wrapper
* Fixing optimizer test after merge

* init cudnn batch norm op

* rename batch_norm_cudnn_op.cc batch_norm_op.cu

* correct name style

* add ExtractNCWHD, simplify code

* fix ExtractNCWHD

* use CUDNN_ENFORCE instead of PADDLE_ENFORCE

Fix new remote updater sending all 0 param / grad to pserver

Fix new pserver having random momemtum initial values

fix package name bug