Add a variable to control whether to retain the graph or not in the backward pass

[BenCrulis]

With the addition of this variable, you can decide to keep the graph by setting strategy.retain_graph = True so that plugins can do their own backward pass after to the one triggered automatically without having to execute the forward pass again.

Unsatisfactory alternatives:

• subclass BaseSGDTemplate and change the implementation of its backward() method
• monkey patch the backward() method
• call strategy.forward() each time a second backward pass has to be made in the same iteration

Note that perhaps the name of this variable should be changed to something else.

[coveralls]

Pull Request Test Coverage Report for Build 5079535595

• 3 of 3 (100.0%) changed or added relevant lines in 1 file are covered.
• 32 unchanged lines in 10 files lost coverage.
• Overall coverage increased (+0.1%) to 72.249%

---

Files with Coverage Reduction	New Missed Lines	%
avalanche/benchmarks/datasets/clear/clear.py	1	13.02%
avalanche/benchmarks/scenarios/benchmark_wrapper_utils.py	1	73.33%
tests/training/test_supervised_regression.py	1	79.12%
avalanche/benchmarks/datasets/downloadable_dataset.py	2	30.61%
avalanche/evaluation/metrics/amca.py	2	75.36%
avalanche/evaluation/metrics/gpu_usage.py	2	38.33%
avalanche/evaluation/metrics/forgetting_bwt.py	3	49.36%
avalanche/evaluation/metrics/forward_transfer.py	4	49.18%
avalanche/evaluation/metric_definitions.py	7	75.46%
avalanche/benchmarks/datasets/mini_imagenet/mini_imagenet.py	9	17.19%

Totals
Change from base Build 5065785914:	0.1%
Covered Lines:	15845
Relevant Lines:	21931

---

💛 - Coveralls

[AntonioCarta]

Thanks. I didn't consider the fact that plugin may need to set retain_graph, which makes subclassing useless. I think your solution is ok.