Add feature ConfidentMisclassification

[wfleshman]

This is a criteria for generating high confident misclassifications without conditioning on a target class. Attacks of this kind are useful when evaluating the robustness of adversarial defenses.

[coveralls]

Coverage increased (+0.5%) to 100.0% when pulling aeea3f2 on wfleshman:add-confidence into 6f3a637 on bethgelab:master.

[coveralls]

Coverage decreased (-0.04%) to 99.483% when pulling fcd54e0 on wfleshman:add-confidence into 6f3a637 on bethgelab:master.

[jonasrauber]

Looks good. We still need a test in https://github.com/bethgelab/foolbox/blob/master/foolbox/tests/test_criteria.py and an entry in the list of criteria at the beginning of https://github.com/bethgelab/foolbox/blob/master/foolbox/criteria.py#L14, then we should be good to go 👍

[jonasrauber]

I am not sure why coveralls reported 100 % test coverage, because clicking on details clearly shows that the lines are not covered 🤔