BATON: Aligning Text-to-Audio Model with Human Preference Feedback