Open-Vocabulary Segmentation with Unpaired Mask-Text Supervision