Can we globally optimize cross validation loss in ridge regression