Can we globally optimize cross-validation loss? Quasiconvexity in ridge regression