Aligning Text-to-Image Diffusion Models with Reward Backpropagation