Microsoft’s new VASA-1 AI framework generates super-realistic talking heads that can even sing songs