[ENH] support for probabilistic regressors (skpro) in make_reduction, direct reduction

[fkiraly]

The aim of this PR is to extend make_reduction classes to support regressors with probabilistic prediction modes, with the skpro interface.

This would be achieved exactly as currently in YfromX which already does support probabilistic regressors, but unfortunately the inner structure of make_reduction is much more nested and organically grown.

The idea is to switch over predict and predict_interval internally, and leave the rest as-is.

Hopefully, this will also not interact with the fixes to the reducers that @benHeid is currently working on.

Depends on #5539, #5725, #5726, which should be merged first.

Also fixes an unreported bug where the reducers could fail in a case where not all index levels were named, as said names were used as reference (in _cut_df). This has been remedied by using integer references.

[benHeid]

Looks good to me. Happy to approve the PR, after the one question I have is answered. I would think, that using pre-allocated arrays might improve the performance. However, I am also aware that they are very confusing...