scikit-learn: More efficient MissingIndicator.fit_transform

Currently, we use self.fit(X, y).transform(X). In this way, we actually call _get_missing_features_info twice, which is not necessary.