Self-supervised Monocular Depth and Pose Estimation for Endoscopy with Generative Latent Priors