Created by: jczaja

Changes disscused here are turning pool mkl_dnn (grad) op into implementation using common interface eg. Acquire API

Note: Backward Op used to enforce to work on same formats as forward , with those changes backward does operate on format of its own choice 'any'. In theory this should be more efficient than making additional reorder of data just to match forward data formats. On the other hand if pooling grad is work on 'any' format then it may happen that pooling_backward working in NCHW16C has to understand workspace memory (transfered from forward) that was created with NCHW format.

@luotao1 If you have any training workload that is using pool2d op and you have it running under mkl-dnn then please test those changes against that workload and let us know the result