CronusVLA: Towards Efficient and Robust Manipulation via Multi-Frame Vision-Language-Action Modeling