How to handle X_val, y_val for TransformersEstimator

Currently X_val, y_val are handled differently for TransformersEstimator in [ml.py]. (https://github.com/microsoft/FLAML/blob/main/flaml/ml.py#L584)

The goal is to remove this special treatment for NLP by removing these lines. Allow X_val, y_val to be None.

If X_val, y_val are none, change the code in TransformersEstimator, so that we don't checkpoint the intermediate results. Instead, only evaluate the final result outside of TransformersEstimator,fit at [ml.py](https://github.com/microsoft/FLAML/blob/main/flaml/ml.py#L413).

Add a test where X_val, y_val = None.