Implement sampler-based dataloading logic

[lrzpellegrini]

This PR changes how dataloaders are created in data_loader.py.

The new mechanism:

• Supports the PyTorch distributed sampler, which requires calling set_epoch. This will be useful in the immediate future to merge the support for distributed training in the main branch.
• Will create DataLoaders (+Samplers objects) in the __iter__ method. This is needed to manage stateful samplers. The overall length of the final dataloder is computed in the constructor.
• Default collate functions are taken from the newly created collate_functions.py. Existing collate functions have been moved there. Aliases have been created in data_loader.py (those aliases will probably be deprecated in the future).
  • collate_functions.py contains additional functions used to better manage the broadcast of elements while distributed training. They will be useful later!

I will remove the related changes from the already opened distributed support PR #996, from which they were taken.

[coveralls]

Pull Request Test Coverage Report for Build 2677446091

• 99 of 134 (73.88%) changed or added relevant lines in 2 files are covered.
• 2 unchanged lines in 1 file lost coverage.
• Overall coverage decreased (-0.005%) to 72.241%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/benchmarks/utils/collate_functions.py	14	23	60.87%
avalanche/benchmarks/utils/data_loader.py	85	111	76.58%

Files with Coverage Reduction	New Missed Lines	%
avalanche/benchmarks/utils/data_loader.py	2	81.28%

Totals
Change from base Build 2641471953:	-0.005%
Covered Lines:	12765
Relevant Lines:	17670

---

💛 - Coveralls