Normalized Wasserstein Distance for Mixture Distributions with Applications in Adversarial Learning and Domain Adaptation