Rar merge

[JuliousHurtado]

This is the first version of the implementation of Retrospective Adversarial Replay (RAR) for Continual Learning.

I tested on CIFAR10 using RAR, and compared it with ER (as in the original paper). RAR improves performance over simple ER.

The exact results shown in the paper could not be replicated, mainly due to the fact that the ER shows much higher accuracy and no original code is provided.

Memory size	ER	ER + RAR
200	32.76% + 4.03%	37.62% + 2.46%
500	37.85% + 2.61%	43.43% + 2.40%

[coveralls]

Pull Request Test Coverage Report for Build 5409909008

• 20 of 102 (19.61%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.2%) to 72.37%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/training/plugins/rar.py	19	101	18.81%

Totals
Change from base Build 5399888339:	-0.2%
Covered Lines:	16386
Relevant Lines:	22642

---

💛 - Coveralls

[AntonioCarta]

Thanks @JuliousHurtado. Can you fix the syntax error and adjust the doc a little bit?