refactor: don't rely on internal scikit-learn functions

[jolars]

We are current using _preprocess_data from sklearn.linear_model, but it is not a public API and should not be used. This just created a problem for me since they dropped the return_mean argument, which caused the benchmark to fail.

[mathurinm]

It's a concern for me too in celer. I have not found a satisfying solution yet