MotionBERT: A Unified Perspective on Learning Human Motion Representations