Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model