Human-Alignment and Calibration of Inference-Time Uncertainty in Large Language Models