Make explicit that cross_val* functions uses unshuffled KFold CV

Describe the workflow you want to enable

Many algorithms that use cross-validation support supplying a simple integer value to the cv parameter for simplicity. This will then create a KFold or StraifiedKFold with the appropriate number of folds. Many estimators and folding strategies also support the random_state parameter for reproducibility, however this can be broken if your folding strategy does not also have a random_state set.

Describe your proposed solution

Allow estimators to pass their random_state object to check_cv so that if a use sets random_state, everything is completely reproducible if an integer was used for cv.

Describe alternatives you’ve considered, if relevant

Additional context

[NicolasHug]

Perhaps we should extend the current docs (addition in italics):

For integer/None inputs, if the estimator is a classifier and y is either binary or multiclass, StratifiedKFold is used. In all other cases, KFold is used. These splitters are instanciated with shuffle=False so the splits will be the same across calls.

[glemaitre]

This issue has been addressed in #19776