Create common sense unit tests?

[mfeurer]

This is mostly a discussion item. PR #1402 unveiled a bug in the handling of pandas that could lead to drastically worse performance (see here). It was uncovered because the training performance of Auto-sklearn was way below 70% accuracy. Since we cannot test everything, are there common sense unit tests we could employ besides checking the training accuracy and maybe that there are not too many crashes of base algorithms when running Auto-sklearn in the unit tests?