Multi-modal Prototypes for Open-World Semantic Segmentation