Inference on the Best Policies with Many Covariates