Summary:
Add CUDAContext::cudnn_handle() for easier integration of single
cudnn routines into operators without requiring the weight
of CuDNNWrapper or similar, or needing to spin out a separate CuDNN*Op
version of an operator.
It was necessary to split out the cuDNN wrapper code from the base cuDNN helpers in order to resolve a circular dependency between context_gpu.h and common_cudnn.h when handles and cuDNN `#define` were added.
Closes https://github.com/caffe2/caffe2/pull/1376
Reviewed By: pietern
Differential Revision: D6162034
Pulled By: akyrola
fbshipit-source-id: 95687e55b3e1e921e1f5e0f016f43b586f5f3350
