Describe the workflow you want to enable

For LogisticRegression with n_classes >= 3, I would like to specify a class-wise (grouped) L21 penalty $\sum_{j=1}^{n_{features}} \Vert\beta_{j,\cdot}\Vert_2 = \sum_{j=1}^{n_{features}} \sum_{k=1}^{n_{classes}}|\beta_{j,k}|$ instead of the implemented L1 penalty $\sum_{j=1}^{n_{features}} \sum_{k=1}^{n_{classes}}\Vert\beta\Vert_2$.
The difference is bassically the same as calling MultiTaskElasticNet().fit(X, y) vs ElasticNet().fit(X, y.ravel()).

glmnet has the option type.multinomial = "grouped" and (default) type.multinomial = "ungrouped".

Describe your proposed solution

Introduce a new parameter, e.g. l1_type or similar.

Describe alternatives you've considered, if relevant

No response

Additional context

I could not find any discussion in and around of #8446 (SAGA was the first and so far only solver capable of solving multinomial case with L1 penalties).