LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models