RMAdapter: Reconstruction-based Multi-Modal Adapter for Vision-Language Models