Enhancing Adversarial Contrastive Learning via Adversarial Invariant Regularization