Multimodal Image-to-Image Translation via Mutual Information Estimation and Maximization