LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models