Open-Vocabulary Calibration for Vision-Language Models