NaN propagation for GPU pooling ops (#12504)

- Enable custom fwd maxpooling kernel to propagate NaNs. This
  makes it match the behaviour of CUDNN, and ensures that CUDNN's
  bwd maxpooling kernel behaves as expected (propagating NaNs).
- Previous behavior remains default. To enable nan-propagation,
  set TF_ENABLE_MAXPOOL_NANPROP=1.
- GPU bwd maxpool op tests cover both propagated and not-
  propagated NaNs.

Performance is unaffected by change. On P100 GPU:
- //tensorflow/core/ops_nn_ops_test is 12ms before and after
- //tensorflow/python/kernel_tests:pooling_ops_test takes
  84.0 sec before vs. 83.8 sec after change (delta is in the
  noise).