Boosting Demographic Fairness of Face Attribute Classifiers via Latent Adversarial Representations