[metrics] (MeanScores) Add a mean scores metrics inspired from IL2M

[Mathieu4141]

To prepare #676

This is a metric that was used to show bias toward new classes in the IL2M paper.
The scores of the true class are averaged by past and new classes.

• For the train plugin I only used the last epoch, the graphs are hard to read if we add more
• For the eval plugin the full stream needs to be considered, so that we can compare old and new classes

This is what I get for the simple example:

The MeanScores metric can be reused for the IL2MPlugin

Edit: Ready for review :)

[coveralls]

Pull Request Test Coverage Report for Build 969419435

• 1 of 5 (20.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.02%) to 79.095%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/evaluation/metrics/mean.py	1	5	20.0%

Totals
Change from base Build 964673211:	-0.02%
Covered Lines:	9504
Relevant Lines:	12016

---

💛 - Coveralls

[AndreaCossu]

Thanks @Mathieu4141, seems good! Have you considered building your plugin metric on the GenericPluginMetric interface? If it was possible, it would make the metrics more homogenous. Let me know 😄

[Mathieu4141]

Thanks @Mathieu4141, seems good! Have you considered building your plugin metric on the GenericPluginMetric interface? If it was possible, it would make the metrics more homogenous. Let me know 😄

@AndreaCossu I tried at first to do it with GenericPluginMetric. However this plugin doesn't quite fit the template so I switch to a custom plugin, and it simplified things a lot
Some of the issues I had:

• The eval plugin would get reset during eval, so we would loose the new/old classes distinction that can only be taken from the previous training. In a custom plugin I simply get the classes in before_training_exp, and reset at before_training
• The train plugin would work on all epochs, but the graphs are really difficult to read and interpret then, doing it on the last epoch seems better

Those issues are really specific to this plugin, so I didn't want to change the generic class only for that

[AndreaCossu]

Thank you again @Mathieu4141 , seems all good to me, I will merge this 😄