Add FromScratchTrainingPlugin

[HamedHemati]

This PR adds a new plugin called FromScratchTraining that resets the model's weights and optimizer state after each training experience. It is used as a baseline for comparison with sequential fine-tuning in some papers.

Minor: we can change the plugin's name if it's wordy. Another potential name could be: "StrategyResetter"

[coveralls]

Pull Request Test Coverage Report for Build 4024812550

• 13 of 30 (43.33%) changed or added relevant lines in 3 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.05%) to 73.671%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/training/supervised/strategy_wrappers.py	3	8	37.5%
avalanche/training/plugins/from_scratch_training.py	9	21	42.86%

Totals
Change from base Build 3998183234:	-0.05%
Covered Lines:	13800
Relevant Lines:	18732

---

💛 - Coveralls

[AntonioCarta]

I think it's better to use the SGDTemplate here to support more use cases.

[HamedHemati]

Agreed, will change it!

[AntonioCarta]

Should we have a supervised strategy too? Since we have them for the other baseline methods.

[HamedHemati]

Should we have a supervised strategy too? Since we have them for the other baseline methods.

Updates:

• Updated the class definition
• Added strategy wrapper
• Added example

[AntonioCarta]

Thanks!