Convex-adversarial

Notable changes:
+ Removed extraneous arguments that had no meaning (l1_eps, m, k)
+ Changed naming of arguments specifically labeled for L-infinity perturbations to allow for L2 perturbations
+ Added trained models in `models_scaled`. These are all cascaded models for MNIST and CIFAR at the epsilon levels described in the paper.
+ Added code for L2 perturbations in the input space.

This version is on PyPI under version 0.4 and works with PyTorch 0.4.

This version reproduces the results of [Provable defenses against adversarial examples via the convex outer adversarial polytope, by Eric Wong and J. Zico Kolter](https://arxiv.org/abs/1711.00851).