- `GridSearchCV` now does not explicitly raise an error if there was an error in fitting the estimator. Rather, it simply returns `nan` as the score. This is not what we want in SKLL, so we set the `error_score` parameter to `raises` which will behave as expected.

…nto 653-update-sklearn-to-0-24-1

- This requires setting `_use_dense_features` since sklearn requires non-sparse features for this type of model.