DeepSeek-VL2: MoE Vision-Language Models for Advanced Multimodal Understanding