Created by: wojtuss

This patch adds a mkldnn-placement-pass pass which implements a very simple strategy for MKL-DNN op placement.

This patch also refactors conv+bn fuse pass (actually, it includes changes from PR #13926) updated to use the mkldnn-placement-pass. Now, if MKL-DNN is used, the fuse conv+bn gives a conv with bias only, and conv+elementwise_add otherwise.