Copy-Pasting Coherent Depth Regions Improves Contrastive Learning for Urban-Scene Segmentation