MM-Embed: Universal Multimodal Retrieval with Multimodal LLMs