Multi-Frame Video Prediction With Learnable Temporal Motion Encodings