Revamp cudnn_rnn ops level unittests

* Use absl parameterized.named_tuple decorator
* Test training by comparing outputs and grads of cudnn and tf rnn, instead of comparing cudnn's symbolic and numerical gradients. Much more comfortable tolerance levels.
* Split different rnn type's tests, much easier to read. (though paying a little more cost in coding.)

PiperOrigin-RevId: 221502090