Interpretable Preference-based Reinforcement Learning with Tree-Structured Reward Functions