Add protocol for handling X_val, y_val

The handling for X_val, y_val for NLP model needs to be taken care of differently. In NLP mode (and only in NLP), one trial can have multiple checkpoints. If a trial has more than 1 checkpoints, how to evaluate the trial's perf for cross-validation is not straightforward.

The following protocol should be added:

If NLP is not under the cross-validation model: requires X_val, y_val for TransformersEstimator.fit, raise an exception if the input does not contain X_val, y_val.
If cross-validation: only evaluate at the end of the training.