Bridging VLMs and Embodied Intelligence with Deliberate Practice Policy Optimization