UCF101 scenario epsilons are too large

[dxoigmn]

Given the c14n of datasets in Armory 0.12, the attack eps value in the UCF101 scenarios is too large now (51), which means the attacker is always successful. These values should be normalized to [0, 1] range. I imagine a similar analysis should be done for the other scenarios to make sure the eps values are in the correct range.

[davidslater]

Thanks for noting this. I made a fix. Checking the other scenarios, they appear to be properly scaled. Audio is a bit weird, as the scale can vary quite a bit more than images/video, but linf is not a great metric in that space.