Link-Context Learning for Multimodal LLMs