General-purpose validation and model selection when estimating individual treatment effects