An Automatic Finite-Sample Robustness Metric: When Can Dropping a Little Data Make a Big Difference?