lifelong-learning

I noticed it is quite tricky at the moment to generate a benchmark with an unbalanced number of examples for each step.
It would be nice to have an option in ni_scenario, nc_scenario and similar to set the number of examples for each step.

We should create tutorials for most of our main datasets and all scenarios.

We should have a jupyter version, and more importantly a Colab version that users can try our library quickly.