Rework baselines

[alafage]

Current baseline implementations carry lots of repetitions. To simplify them this pull request proposes the following changes:

1. Classification Single/Ensemble routines are no more abstract. They can be used directly with models, losses, etc.
2. Unique baseline for each backbone (ResNet, WideResNet, etc.) where the version (ex: vanilla, packed, etc.) is passed as an argument.

These changes are expected to:

1. reduce considerably the number of files for the baselines
2. ease the use of routines to train custom models

• Routines are not more abstract
• ResNet baseline
• WideResNet baseline
• ResNet experiments
• WideResNet experiments
• Custom optimizer management for BatchEnsemble
• Simplify the arguments (of packed, etc) but also of the routine (metrics)
• Simplify the argument consistency tests
• Tests
• Documentation

[o-laurent]

You also moved loading weights inside the model! One step further to solving #11

[alafage]

I think it should be:

if model_name == "batched":
    ...

[o-laurent]

You're right. However, is batched the correct name? Calling BatchEnsemble "batched" may amount to some form of appropriation. (says the guy that called BatchEnsemble "BatchEnsembles" twice in the commits 😆)

[o-laurent]

I fixed the optimizer, but we may need to look at the names we gave to the different methods.

[o-laurent]

Starting to look nice!