Add unpenalized logistic regression

The problem sets of logistic regression all have a penalty. It would be very interesting, at least to me, to add the zero penalty case.

Note: For n_features > n_samples, like the 20 news dataset, this is real fun (from an optimization point of view).

Start by locating the benchmark's logistic-regression implementation and the configuration or problem definitions for the existing penalty cases. Trace how the penalty is passed into the optimization setup, then determine how a zero-penalty case should be represented and validated, especially when n_features exceeds n_samples. Done means the unpenalized case is included consistently with the existing benchmark cases and its results can be checked.

Written by the indexing model from the issue text.