Created by: Sand3r-

This PR fixes #20842 (closed). The manual fallback to native Paddle kernel is removed, now that happens through GetExpetectedKernelType mechanism. That means: GetExpectedKernelType of ElementwiseMul checks if:

• x input is of format nChw16c
• y input is of format nc
• channels are divisable by 16 And only all those conditions are true, returns an MKL-DNN kernel. Otherwise the native Paddle kernel is returned. If that happens, all the necessary tensor format conversions that were taking place in the kernel are now delegated to the code in data_layout_transform.cc

Also it removes x_data_format and y_data_format attributes resolving #16948 (closed).