AlignGAN: Learning to Align Cross-Domain Images with Conditional Generative Adversarial Networks