InterVLS: Interactive Model Understanding and Improvement with Vision-Language Surrogates