Regularization properties of adversarially-trained linear regression