Representation Learning via Adversarially-Contrastive Optimal Transport