DRAGON: Distributional Rewards Optimize Diffusion Generative Models