Raise for unseen categories when materializing from an existing ModelSpec

[stanmart]

Checklist

• Added a CHANGELOG.rst entry

Added a check for unseen levels when materializing a matrix from an existing modelspec. This is analogous to raising an error when trying to predict from categoricals having levels unseen in the training data. This error is especially useful for glum, because it solves glum PR #748 when glum with formulas.

It also makes tabmat's formula materialization stricter than formulaic's. The latter only emits a warning in this case, and sets all dummies to zero. I think the proposed tabmat behavior also makes sense, especially in conjunction with glum, but let me know if you think otherwise. (Maybe one point for discussion could be what should happen if cat_missing_method="zero". In the proposed PR, unseen categories still raise a ValueError, but a warning plus setting all dummies to zero might also make sense in that case. Same question goes for glum's behavior.)

[MatthiasSchmidtblaicherQC]

Thanks! Looks great, just some minor comments.

I am for raising on unseen categories also under cat_missing_method=="zero", i.e. the current proposal. After all, cat_missing_method is about missings, so it would be strange if it also governed behavior for other categories. Going with the more restricted version first is also easier because, if we decide to change it later, it won't be a breaking change.

[MarcAntoineSchmidtQC]

Looks good. Thanks @stanmart ! Happy to merge once @MatthiasSchmidtblaicherQC's comments have been addressed.