Benchmarking Robustness of Adaptation Methods on Pre-trained Vision-Language Models