Robust extrapolation designs and weights for biased regression models with heteroscedastic errors