Tests failing with formulaic>=1.1 and scikit-learn>= 1.6

[e-dorigatti]

Tests on main fail since a few days, and involve multiple recent PRs.

Culprit seems to be a new release of a dependency that breaks Jacobian computation. I've already opened an issue there matthewwardrop/formulaic#235

Scikit learn 1.6 is also changing a few things that break xgboost 2.1.3, presumably only until their next release as the required changes were already made dmlc/xgboost#11021