Risk bounds in linear regression through PAC-Bayesian truncation