MKL support for max/avg pooling and relu (#8296)

* Adding MKL support for Max/Avg Pooling and ReLU

* Missed the mkl layer registry files

* Fixed sanity check errors with buildifier

* Adding MKL support for Max/Avg Pooling and ReLU

* Missed the mkl layer registry files

* Fixed sanity check errors with buildifier

* Adding Intel Conv2D kernel implementation alongwith required Graph passes

This commit contains 4 main components:

1) Intel-optimized kernel implementation for Conv2D op
   Implementation in kernels/mkl_conv_ops.*

2) Graph passes required to enable Conv2D optimized implementation
   Implementation in graph/mkl_*. We also need a new op, MklToTf op.
   Its implementation is in kernels/mkl_tfconv_op.cc.

3) Utility functions used in kernel implementation
   Implementation is in common_runtime/mkl_layer_registry* and util/mkl_util.h

4) BUILD changes for Conv2D, graph passes and utility functions

* Refactor MKL convolution forward pass computation into smaller functions. Changed configure to point to newer MKLML library

* Moved Mkl helper datastructures and routines to private class members

* MKL op registration changed to use existing op registry (nhasabni)

* Fixed buildifier error

* Adding MKL support for Max/Avg Pooling and ReLU

* Missed the mkl layer registry files

* Fixed sanity check errors with buildifier

* Removed the mkl layer registry (should not have been added) and made fixes according to the code review comments

* Adding Intel Conv2D kernel implementation alongwith required Graph passes

This commit contains 4 main components:

1) Intel-optimized kernel implementation for Conv2D op
   Implementation in kernels/mkl_conv_ops.*

2) Graph passes required to enable Conv2D optimized implementation
   Implementation in graph/mkl_*. We also need a new op, MklToTf op.
   Its implementation is in kernels/mkl_tfconv_op.cc.

3) Utility functions used in kernel implementation
   Implementation is in common_runtime/mkl_layer_registry* and util/mkl_util.h

4) BUILD changes for Conv2D, graph passes and utility functions

* Refactor MKL convolution forward pass computation into smaller functions. Changed configure to point to newer MKLML library

* Moved Mkl helper datastructures and routines to private class members

* MKL op registration changed to use existing op registry (nhasabni)

* Fixed buildifier error

* Adding MKL support for Max/Avg Pooling and ReLU

* Missed the mkl layer registry files

* Fixed sanity check errors with buildifier

* Removed the mkl layer registry (should not have been added) and made fixes according to the code review comments

* Fixed rebase messups

* Added documentation for mkl pooling op parameters

* removed layer registry reference from mkl relu op