On the Effectiveness of Least Squares Generative Adversarial Networks