STAR: SpatioTemporal Adaptive Reward Allocation for Text-to-Image RL Post-Training