Velodrome combines semi-supervised learning and out-of-distribution generalization (domain generalization) for drug response prediction and pharmacogenomics