Text-to-Image Diffusion Models are Zero-Shot Classifiers