Reproducibility & new methods, settings, and settings

[loribonna]

Introduction of REPRODUCIBILITY.md, to keep track of which methods we have already a configuration available that reproduce the results of the original papers. Several methods have been checked, with small refactors and updates to support the setup used in their original works.

For example, for iCaRL we reproduced the original ResNet32 with the identity shortcut.

To reproduce the results of a method, check if it is supported in REPRODUCIBILITY.md and then run:
python main.py --model <model_name> --dataset <dataset_name> --model_config best [--buffer_size <buffer_size>].
For more information, visit the documentation.

Included new datasets and models:

• added support for datasets with bias (seq-celeba)
• added Learning without Shortcut (LwS)
• created several configurations for datasets and models

Improved the documentations with new info and some tutorials.