FIX scoring != None for RidgeCV should used unscaled y for evaluation

[glemaitre]

closes #13998
closes #15648

While discussing with @jeromedockes, we recall to have observed something weird in the RidgeCV code. I check a bit closer and I open this PR to highlight what is the potential problem.

In RidgeCV, when having sample_weight we scale the data using the sqrt(sample_weight):

scikit-learn/sklearn/linear_model/_ridge.py

Lines 2133 to 2136 in 35164b3

	if sample_weight is not None:
	X, y, sqrt_sw = _rescale_data(X, y, sample_weight)
	else:
	sqrt_sw = np.ones(n_samples, dtype=X.dtype)

The idea is that the mean squared error can be expressed as:

scikit-learn/sklearn/linear_model/_base.py

Lines 212 to 223 in 35164b3

	For many linear models, this enables easy support for sample_weight because
	(y - X w)' S (y - X w)
	with S = diag(sample_weight) becomes
	||y_rescaled - X_rescaled w||_2^2
	when setting
	y_rescaled = sqrt(S) y
	X_rescaled = sqrt(S) X

Those "centered" data are used to optimize the ridge loss. Later in the code, we want to compute a score that can be an arbitrary metric via a scorer.

scikit-learn/sklearn/linear_model/_ridge.py

Lines 2158 to 2169 in 35164b3

	predictions = y - (c / G_inverse_diag)
	if self.store_cv_results:
	self.cv_results_[:, i] = predictions.ravel()
	score_params = score_params or {}
	alpha_score = self._score(
	predictions=predictions,
	y=y,
	n_y=n_y,
	scorer=scorer,
	score_params=score_params,
	)

The problem here is that predictions is computed efficiently as provided in the GCV paper. But these predictions are in the "scaled" space and it seems incorrect to compute any metric in this space with an arbitrary metric. Instead, we should unscale these predictions and the scaled true targets to compute the metric in the original space.

This is what this PR is intended to. I did not add any non-regression test (I assume that using the MedAE should lead to some failures) because I wanted to be sure that what I'm saying is correct.

@jeromedockes @ogrisel @lorentzenchr Does the above description make sense to you?

Edit: It seems that it relates to #13998 and #15648

Probably, I should check the tests that were written in #15648

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: e9bd778. Link to the linter CI: here}

[ogrisel]

Cross-linking #16298 as it might be related.

[glemaitre]

So I added a test that was failing on main but is passing in this PR.
So we should be better. However, I discovered a new bug when dealing with sample_weight when scoring != None and with several target.

I'm going to open another PR to not overload this PR.

[glemaitre]

I added a new parametrization to check that we support multioutput properly.

[lorentzenchr]

@glemaitre Could you fix the typos in the whatsnew entry?