Discffusion: Discriminative Diffusion Models as Few-shot Vision and Language Learners