CUT: Contrastive Learning for Unpaired Image-to-Image Translation