Contextual Bandits and Imitation Learning with Preference-Based Active Queries