ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching