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.

and remove some comments.

add the softsign activation.

fix bug (conv2d  groups: 'num_channels % groups is not 0:' )

enable mkldnn benchmark with googlenet

Fix pool max with index.(Mask type should be int, not float)

Add cudnn pool3d

Add sequence slice operator

remove redundant tests in layer helper's unittest.

Add the configuration helper for multi-head attention.