Sample-Optimal Agnostic Boosting with Unlabeled Data