[FEATURE]: Szymanski Sampling

[JacksonBurns]

Is your feature request related to a problem? Please describe.

In the domain of multilabel-classification, data must be partitioned s.t. labels and label pairs are balanced between train, val, and test (see this paper)

Use-cases/examples of this new feature

This has been implemented in scikit-multilearn but unfortunately mostly left abandoned. Would be nice to pick up and include here.

Desired solution/workflow

Could use their reference implementation, as well as the the original implementation which would need to be ported to Python 3.

Discussion

Stepping into classification contexts would be a new direction for astartes, but an interesting one.