ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model