Support more metrics in PHE

[LeoGrin]

Some discussion in #12
Some metrics: balanced accuracy, balanced log loss, brier score....
Some points to consider: maybe for some metrics we want to use balance_probabilities or add it to the search space. How well do we do on the available metrics? We also probably want to pass the metric the the TabPFN model itself.

[LennartPurucker]

FYI: for supporting almost everything sklearn supports, I can recommend using AutoGluon["core"] to get its metric handling or vendoring it like I did for fine-tuning.