Set drop NaN component in time series pipeline to be training-only

Currently, we run the drop NaN component in time series pipelines during fit, transform, and predict. However, we only need to drop NaN values during training. When passing a test set though the pipeline, we shouldn't generate NaN values with the featurizer and thus do not need to have NaN rows to be dropped.

This has implications for calculating features for time series to be used for permutation importance.