Created by: jianhang-liu

When data flow from a MKLDNN OP kernel into a non-MKLDNN OP kernel, data layout transform will be done by framework automatically (see data_transform.cc & data_layout_transform.cc) via doing a MKLDNN reorder. However, when those two OP kernels share the same data layout (or "memory format" more accurately), this reorder should NOT be done.

@tensor-tang @tpatejko @kbinias Please help to review. Thanks!