On the marginal likelihood and cross-validation