Multivariate linear regression with non-normal errors: a solution based on mixture models