Created by: Sand3r-

Purpose

PR updates MKLDNN to version 0.15 and corrects current integration of convolution so that it doesn't crash whenever new padded format is used.

Changes introduced

The above has imposed a requirement that the size of output tensor used by MKLDNN can no longer be assumed to have N * C * H * W * sizeof(DataType). Now the size of the output tensor is determined by mkldnn's engine and queried through a get_size() method. That required modification of mutable_data() method of Tensor.