[MKL-DNN] Reduce memory footprint by using custom scratchpad of MKL-DNN 1.0
Created by: jczaja
Once MKL-DNN 1.0 is merged it will be possible to use custom (allocated by Paddle) scratchpads (temporary buffers used internally by MKL-DNN). This way we possibly reduce peak memory consumption as in 0.20 scrathchpad are allocated for each op separately which is not always needed for all PaddlePaddle usecases.