Created by: jczaja

This is first of series of PRs on tensor modifications for MKL-DNN operations that are replacing mkl-dnn format storing with MKL-DNN primitive descriptor storing.

Idea is that currently MKL-DNN format is stored in Tensor. format is not good enough for Transpose MKL-DNN as format is ambigious in terms of idenditifcations of actual mkldnn-layout. More over it does require recreation of primitive descriptions for given tensor.

Adventages:

• Design should be clearer
• Redundant recreation of user primitive desc should be removed

In this PR:

• Tensor was modified to store MKLDNN primitive desc
• Transpose op was adjusted to work with new functions
• ConvGrad MKL-DNN was barely adjusted as well as Gaussian Random mkl-dnn.
• All other mkl-dnn ops were not modified . They will be adjusted in next PRs (convolution will come next)
• No changes to non MKL-DNN PaddlePaddle (core PaddlePaddle concepts are not modified)