T-Person-GAN: Text-to-Person Image Generation with Identity-Consistency and Manifold Mix-Up