From my understanding:

best_estimator_ provides the estimator with highest score; best_score_ provides the score of the selected estimator; cv_results_ may be exploited to get the scores of all estimators.

However, it is not clear to me how to get the estimators themselves.

As I see it, you cannot. But what you can do is taking the best parameter combination from best_params_ and fit the model again with those same parameters. Check out attributes of GridSearchCV