Uni-MoE: Scaling Unified Multimodal LLMs with Mixture of Experts