An Automatic Finite-Sample Robustness Metric: Can Dropping a Little Data Change Conclusions?