Task Vectors [in Vision and Language Models (VLM's)] Are Cross-Modal