Avoid Wasted Annotation Costs in Open-set Active Learning with Pre-trained Vision-Language Model