Created by: jczaja

Changes disscused here are removing transfering of Tensor X from FWD activation op to backward activation op via device context (this was not working properly for parallel executor based trainings). Instead for mkl-dnn activations Tensor X is an input , hence we take needed info from X tensor directly : x->format(), x_data and we recreate mkl-dnn memory object : src_memory based on those two transfered quantities.