STAU: A SpatioTemporal-Aware Unit for Video Prediction and Beyond