UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent