Scale-constrained approaches for maximum likelihood estimation and model selection of clusterwise linear regression models