Revisiting Superpixels for Active Learning in Semantic Segmentation with Realistic Annotation Costs