Uncertainty in Language Models: Assessment through Rank-Calibration