`D0CRT` Accept Scikit-learn models directly instead of parameter dictionaries

[jpaillard]

Working on #253, I realize that a lot of code would need to be rewritten to adapt to the classification case, while not much is changing in theory.
One of the reasons for that is the use of dictionaries of parameters (like params_lasso_distillation_x) and the method: _fit_lasso, which are incompatible with LogisticRegression.

I suggest replacing these dictionaries with Scikit-learn models.

[lionelkusch]

This is quite similar to the issue #325.
As mentioned before, the usage of more abstract models can create difficulty for the computation of sigma2:

hidimstat/src/hidimstat/distilled_conditional_randomization_test.py

Line 583 in 3ecce71

sigma2 = np.linalg.norm(X_residual) ** 2 / n_samples + alpha * np.linalg.norm(

The computation of sigma2 is based on alpha and the coefficient of lasso. This is very restrictive for compatibility with the Scikitlearn Model.

[jpaillard]

sigma2 is for the regression case; this line is not needed for the classification case.

We could add a check to make sure that the estimator is compatible with the problem type?

[lionelkusch]

You need to change the statistical test; there is a need to include a parameter for defining the type of problem and checking that the estimator is correctly associated with it.

[jpaillard]

If there is consensus that it makes sense to have this kind of test, I will open a preliminary PR for this issue before tackling #253

[lionelkusch]

I agree that starting by removing the dependence of D0CRT on Lasso before including the possibility of classification is the best way of doing it.

[bthirion]

In general, we should resort on sklearn's estimators in every place where it's possible; Otherwise, we will have to maintain a lot of duplicate code that could be avoided. The variance estimation problem can be tackled by proper means.