Add cudnn7 support (#12503)

Fix cudnnSetRNNDescriptor API change
Add support for CUDNN_TENSOR_OP_MATH in cudnn v7
- Applies to forward and backward convolutions that have fp16
  input/output. Computations will fall back to pseudo-fp16
  if tensor op math is disabled or not supported.
- Enabled by default, but can be disabled using the environment
  variable TF_DISABLE_TENSOR_OP_MATH=1.
- The choice of whether to use tensor op math is included in the
  autotuning of convolutions.