Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation