UniT: Multimodal Multitask Learning with a Unified Transformer