Ramen: Robust Test-Time Adaptation of Vision-Language Models with Active Sample Selection