Does Monocular Depth Estimation Provide Better Pre-training Than Classification for Semantic Segmentation?