Merge develop back into master

[andrewilyas]

Probably enough for version 1.2 of the library?

• Fixed random_perturb for robust models
• Add random_step (useful for training randomized smoothing models)
• Mixed precision support but without resuming
• Fix #37 (with_image option in AttackerModel)
• Allow for epsilon schedules (removed the eps-fadein argument in favor of a more general epsilon multiplier schedule)
• Allow resuming optimizer instead of just model with --resume-optimizer (Fix #41)
• make_and_restore_model now makes the model parallel even if resume_path is not specified, new add_custom_forward option for better interfacing with custom architectures
• New dataset (Places365)
• Improved learning rate scheduling

[andrewilyas]

Ok so right off the bat a couple of things we should fix before we merge:

• ~Right now in make_and_restore_model, parallel=True is on by default, but train expects a model that is not DataParallel, we should fix this (one solution is to make the default value of parallel to be (resume_path is not None))~
• ~We should consider evaluating the model straightaway upon resume to get good values of prec1 that are unaffected by the first epoch of training~
• We should make sure all of the new additions are sufficiently documented (specifically a short addition to the training doc RE mixed precision would probably be useful)

[andrewilyas]

Ok the only immediately missing things before I can merge this in:

• [x] I'm not fully convinced of the best_prec1 thing---could see some problems arising if someone is training adversarially and then wants to resume but with standard training. I think it should still be named either adv_prec1 or nat_prec1 (or something like that, and then we should just have a check for whether it's in the checkpoint).

• [x] Need to update the docs for CLI training (everything else should be basically documented)