When large samples act small: Cluster-robust variance estimation and hypothesis testing with few clusters