Out-of-sample scoring and automatic selection of causal estimators