Contributes to #11865.

Fitted Attributes

After the removal of multi_class and any OvR-logic in LogisticRegressionCV in #32073, there are a few fitted attributes that have now (or always had) a strange data format (I neglect l1_ratios in the following for ease of reading):

• coefs_paths_ is a dictionary with class labels as keys and arrays of shape (n_folds, n_cs, n_features) or similar as values.
  As coef_ is an array of shape (n_classes, n_features), coefs_paths_ should be an array of shape (n_folds, n_cs, n_classes, n_features), such that coefs_paths_[idx_fold, idx_cs, :, :] gives comparable coefficients. Maybe the intercept should be separated as intercepts_paths_.
• scores_ is a dictionary with class labels as keys and arrays of shape (n_folds, n_cs) or similar as values. All values are the same regardless of the key (class label). This is a relict from OvR.
  A good value would be just an array of shape (n_folds, n_cs)
• C_ is an array of shape (n_classes)
  As the different penalties for classes are gone with the removal of OvR, C_ should be a single float: the single best penalty parameter.
• l1_ratio_ same as C_
• n_iter_ is an array of shape (1, n_folds, n_cs) or similar
  The first dimension should be removed, i.e. shape (n_folds, n_cs)

Deprecation strategy

It is unclear to me how to accomplish the above. Options:

1. Deprecate old attributes and introduce new ones with new names. (time = 2 releases)
2. Same as 1. but then deprecate new ones and reintroduce the old names. (time = 4 releases)
3. Deprecate old attributes and switch behavior in after the deprecation cycle (time = 2 releases)
4. Another option?

Usually, we avoided deprecations options like 3.
@scikit-learn/core-devs recall for comments