MoE-LLaVA: Advancing Sparse LVLMs for Improved Efficiency