Fix integration with IncrementalClassifier

[JuliousHurtado]

When testing the RAR implementation using a model with IncrementalClassifier class, @AlbinSou found a new error.

The problem arises between an incompatibility between IncremetnalClassifier and the feature_extraction functions from pytorch. Because of this, I changed the mask function in the IncrementalClassifier class to be an in_place operation. This solves the first problem.

A second problem was a KeyError problem with optimizer.state_dict(). Because of this, I simplified the RAR implementation. This can help with the performance but sometimes take more time to run.

[coveralls]

Pull Request Test Coverage Report for Build 5461914572

• 2 of 3 (66.67%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.02%) to 72.392%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/training/plugins/rar.py	0	1	0.0%

Totals
Change from base Build 5414134002:	0.02%
Covered Lines:	16388
Relevant Lines:	22638

---

💛 - Coveralls

[AntonioCarta]

Is the in-place operation compatible with backpropagation? Can you use the not in-place version of the operation instead?

[JuliousHurtado]

I don't know if it is compatible con back-propagation, I don't see why not.
But yes, I change it. I run some small experiments with both versions and I don't see any difference in the performance.

Something that Albin suggest was to change the RAR plugins to something similar to ICarl, and having the feature extractor and classifier independently. This would help us remove the name_ext_layer parameters, since we will have a feature extractor that can return the vector. Don't know what you think?

[AntonioCarta]

Having FE and classifier in separate attributes is always convenient if possible, but I don't fully understand what's the issue with RAR. If it's working I think we can leave it as is.