Transfusion: Predict the next token and diffuse images with one multimodal model