Mix-and-Match Tuning for Self-Supervised Semantic Segmentation