Robust Two-Step Confidence Sets, and the trouble with the First Stage F-statistic