VATT: Transformers for Multimodal Self-Supervised Learning from Raw Video, Audio and Text Columbia University Google Cornell University