Output mask for IncrementalClassifier and MultiHeadClassifier

[AntonioCarta]

I needed a couple of changes to the dynamic modules.

two changes:

• DynamicModules adaptation now uses experience instead of dataset. This should make things faster and support other benchmarks (RL, ex-model, OCL) where we may not have a dataset.
• output masking: some units of incremental/multihead classifiers are unused (e.g. if an experience contains classes 2 and 7, 8 units are created, 6 of which end up unused but trained anyway). Masking avoids training those units. Enabled by default but can be disabled.

[ContinualAI-bot]

Oh no! It seems there are some PEP8 errors! 😕
Don't worry, you can fix them! 💪
Here's a report about the errors and where you can find them:

tests/test_models.py:215:9: E306 expected 1 blank line before a nested definition, found 0
1       E306 expected 1 blank line before a nested definition, found 0

[coveralls]

Pull Request Test Coverage Report for Build 2353262781

• 130 of 132 (98.48%) changed or added relevant lines in 10 files are covered.
• 79 unchanged lines in 5 files lost coverage.
• Overall coverage decreased (-0.3%) to 70.251%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/models/dynamic_modules.py	69	70	98.57%
avalanche/training/templates/online_supervised.py	0	1	0.0%

Files with Coverage Reduction	New Missed Lines	%
avalanche/benchmarks/generators/benchmark_generators.py	1	91.81%
avalanche/training/templates/base.py	1	92.08%
avalanche/training/templates/base_sgd.py	1	94.94%
avalanche/training/plugins/lr_scheduling.py	13	83.33%
tests/training/test_plugins.py	63	82.34%

Totals
Change from base Build 2344910382:	-0.3%
Covered Lines:	12048
Relevant Lines:	17150

---

💛 - Coveralls