LLaVA-CMoE: Towards Continual Mixture of Experts for Large Vision-Language Models