Allow prob-based metrics to work

[pseudotensor]

What does this PR implement/fix? Explain your changes.

Handling probabilities with some metrics

How should this PR be tested?

code here: https://github.com/openml/openml-python/issues/1139

Any other comments?

[PGijsbers]

Thanks for the contribution 🎉 I'll try to have a closer look next week. In the meantime, please have a look at the pull request checklist and fill in the missing gaps, such as:

• add the new parameter to the docstring
• create unit tests, in particular also ones that would fail before this PR but succeed after
• add a note to the progress log

[PGijsbers]

Closing this because it is stale. In another rework we will probably base it on scikit-learn scorer objects instead of the metric function so that we have access to whether it needs proba directly instead of as an additional argument.