Bayesian and regularization approaches to multivariable linear system identification: the role of rank penalties