Describir: Diffusion Models for Open-Vocabulary Segmentation