UMFC: Unsupervised Multi-Domain Feature Calibration for Vision-Language Models