Rectifying Shortcut Behaviors in Preference-based Reward Learning