Add mkldnn layouts to Maxpool and Maxpoolbackprop

* Refactored unary elementwise ops into a single interface
that is adaptable to elementwise ops with arbitrary number of inputs.

* Renamed EmitUnaryElementwise -> EmitElementwise.
Implemented first binary elementwise op (Power).

* Refactored some of the boiler plate code for emitting cuda kernels to nvrtc
out of the emit functions and into the CudaFunctionPool static singleton.
CodeWriter now saves cuda kernels to ./gpu_codegen.

* Added ops Divide, Subtract & Sign to the GPU transformer.
Subtract and Sign both use custom device helper functions which
have math kernels defined for the op in gpu_cuda_kernel_ops.hpp,
and which are built by a new get_device_helper function.

NGMX-296 Convolution + Bias with MKLDNN

Also fixed conv+bias cpu layout bugs.

* fixed memory leak.

* clang format.

* update gpu_emitter use template

* add template

* removing extra copies due to op::Result

* remove comment

* fix comment

* switch to a flag version

* add copyright header #pragma once

* add impl file, rename result_elimination.hpp to result_copy_elimination.hpp to match the opt name

* add cpp suffix to result_copy_elimination

* use member in-class member init

* - Added sigmoid fusion pass
- added mkldnn emitter code for sigmoid

* - corrected sigmoid expected values
- add layout assignment for sigmoid op

* - added assert's in cpu fusion for sigmoid
- style fix

* remove debug prints

* NGMX-371 #comment addressed PR comments - Added sigmoid unit test case with 3D input ii) support in cpu_emmiter for sigmoid to handle all input shapes

* NGMX-371 #comment use shape_size() to calculate the 1d input size

Jbobba/batchnorm layouts

* remove broadcast from matmulbias

* fix comments

* working gemm-based broadcast

* fix clang warning

* straightforward gpu.cos implementation following previous patterns prior to refactor

* Generalized unary elementwise gpu op impl.. New unary elementwise ops can
be added to the type annotations in gpu_cuda_kernel_ops.hpp. Next step
is to refactor the llvm interface in gpu_emitters.hpp for similar generality.

* Added gpu_emitter.hpp:EmitUnaryElementwise.

Function adds cuda kernel based on ngraph::op::op_type::description.
This can service all unary elementwise ops run on the gpu.

* The following elementwise unary ops now use the EmitUnaryElementwise emitter:
* GPU.abs
* GPU.acos
* GPU.asin
* GPU.atan
* GPU.ceiling
* GPU.cos
* GPU.cosh
* GPU.exp
* GPU.floor
* GPU.log
* GPU.not
* GPU.sign
* GPU.sin
* GPU.sinh
* GPU.tan
* GPU.tanh
Unary elementwise ops Sign and Not need extra consideration.

* tanh test changed to test::all_close for fp comparison (also done for tan in commit 65fa7c6de34c8277fe2a4801644f6bb64574f4ff).

* GPU backend skips added for recent softmax test and updated aliased output test that uses op::Constant.

* code format update

* changed cuda builder interface names to unary/binary/arbitrary, added impl. note to gpu_cuda_kernel_ops, cleaned code format

* updated ngraph-cpp reference

* Fixing incorrect github conflict resolution.

* Added GPU emitter for op::Result.
For now it simply copies the output tensor.

All but 3 tests now pass. The remaining
failing tests are:
* GPU.dot_0_0
* GPU.dot_matrix_2x0_0x2
* GPU.dot_2x0_0

* Removed call to handle memory aliasing in gpu_external_function.

* fix gpu emitter bug that will return in the middle of function

* Merge pull request #609 from NervanaSystems/tfl/fix_return_bug

fix gpu emitter bug that will return in the middle of function

* GPU backend skips added for recent softmax test and updated aliased output test that uses op::Constant.

Fix constant bug on GPU

* Added GPU emitter for op::Result.
For now it simply copies the output tensor.

All but 3 tests now pass. The remaining
failing tests are:
* GPU.dot_0_0
* GPU.dot_matrix_2x0_0x2
* GPU.dot_2x0_0

* Removed call to handle memory aliasing in gpu_external_function.

* fix gpu emitter bug that will return in the middle of function

* Merge pull request #609 from NervanaSystems/tfl/fix_return_bug

fix gpu emitter bug that will return in the middle of function

* GPU backend skips added for recent softmax test and updated aliased output test that uses op::Constant.

* - Added support optimized bn mkldnn implementation in cpu emitter
- modified bn unit_test to support new implementation
- added layout assignment for bn op
- Style Fix

(cherry picked from commit 7747a40806d62c126059d5c873adcd2e61a0adb0)

* modified value initilization in cpu_fusion to be float explicit

(cherry picked from commit 03499d380073d0197ab8cbd154eb03f63b042a48)

* fix compilation issue

* Addressed PR comments
- added exception if gamma and beta layout isnot equal to memory::format::x
- throw exception if bn Op is not mkldnn op

* fix compilation issue

* added support to handle multiple o/ps in fprop bn fusion

* - Removed laytout pass for bn
- fixed autodiff bug in bn
- added "Add" for the dispatcher in cpu-layout pass

* style fix

* Fix bprop batchnorm test with get_output_elements

* Style fix

- Pass mkldnn workspaces through runtime context
- Move maxpool ops and avgpool backprop to mkldnn emitter