Efficient Meta Reinforcement Learning for Preference-based Fast Adaptation