* add onehot op

* refactor broadcast and onehot op

* Added corresponding cudaFree to the cudaMalloc for the cuda pool_base_ptr memory buffer.

* Check for temporary buffer allocation prior to freeing. Add null check on cudaFree.

* Add cpio file read/write class and unit tests

add reserializer

Add unit test for serialize constants to cpio file. Fix bug in serializer if function has no parameters.

Jbobba/mkldnn v0.13

* GPU elementwise emitters now respect input and output tensor types.
This enables the use of binary comparison ops and op::Convert.

* Removed comments.

* All kernels now have type signature
even if the i/o tensors are equivalent type so that
kernels for specific type tensors are unique.

NGMX-391 #close 

* Fix bn construtor
    - assert if gamma or beta dont have rank 1
    - remove redundant checks

* - added gaurds to check if the input and delta shape to mkldnn bn fprop and bprop op has a rank of 4

gpu dot bug fix for bprop

* Updated namespace use in cpp files.

* - Added pattern matcher for bprop sigmoid
- mkldnn emitter code for sigmoid bprop
- Fusion pass unit test for sigmoid bprop
- style fix

* Added test case for bprop sigmoid

* fixed sigmoid bprop test case failure

* fixed bprop unit test values for sigmoid

* style fix

* fix typo

* Addressed PR comments
- added layout assignment pass to ensure delta and input have same layout for SigmoidBprop

Batchnorm bprop layouts and move the last few mkldnn ops to mkldnn_emitter

* fix detailed timing flag

* more detailed info

use op::Constant's data rather than emitting the data in the generated cpp code. This make compile times for trained models something like 100x faster. (#624)

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.