Uncovering Prototypical Knowledge for Weakly Open-Vocabulary Semantic Segmentation