Pretrained deep models outperform GBDTs in Learning-To-Rank under label scarcity