GME: Improving Universal Multimodal Retrieval by Multimodal LLMs