Mitigating Political Bias in Language Models Through Reinforced Calibration