Created by: Sand3r-

Fuses Convolution with ReLU using transpiler when MKLDNN is used.

Also, due to small re-arrangement in transpiler methods execution order, now the whole [Conv] -> [BN] -> [ReLU] is fused into a single [Conv].

On ResNet-50, this change speeds up the whole topology on OMP_NUM_THREADS=1 and batch_size=1 by 17%. and 33% on CRNN-CTC.