Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation