Joint Adversarial Learning for Cross-domain Fair Classification