On the Entropy Calibration of Language Models