The Surprising Effectiveness of Canonical Knowledge Distillation for Semantic Segmentation