Lp-optimality for regression designs with heteroscedastic errors