Training-Free Open-Vocabulary Segmentation with Offline Diffusion-Augmented Prototype Generation