A Simple and Effective Reinforcement Learning Method for Text-to-Image Diffusion Fine-tuning