Inference on the best policies with many covariates