Add stratify split support for multilabel

[LexABzH]

Problem

We provide a stratify split function, but on a single column.
In some multilabel problems, we would like to stratify our dataset on several columns.

Let's say I have 30% 1s for label A, and 60% 1s for label B. I would like to split my dataset on several sets, and still have these proportions on each new sets. Today I only can apply stratification on a single label.

Concerned template

• NLP template
• NUM template
• VISION template
• API template
• How templates are generated - Jinja

Solution

Adapt stratified_split function to allow serveral columns to stratify on.

Additional context

Also update scripts 0_split_train_valid_test.py.