Aligning Text-to-Image Models using Human Feedback