Text this: Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping