Scene Text Segmentation with Text-Focused Transformers