[MRG] Support for Out-of-Bag cross validation: Add IdentitySplitter() and oob_score()

[multimeric]

Reference Issues/PRs

Fixes #23382

What does this implement/fix? Explain your changes.

• Add IdentitySplitter(), which is a dummy cross-validation splitter that just returns the training data. This can be used to implement cross validation for OOB
• Add oob_score(), a scorer function that delegates to the RandomForest oob score
• Added general tests for both, and for the integration of the two together in a CV context

Any other comments?

The intended usage of these new objects together is as follows:

from sklearn.metrics import oob_score
from sklearn.model_selection import GridSearchCV, IdentitySplitter
from sklearn.ensemble import RandomForestClassifier

cv = GridSearchCV(
    RandomForestClassifier(oob_score=True, random_state=0),
    {"n_estimators": [1, 20, 100]},
    cv=IdentitySplitter(),
    scoring=oob_score,
)
results = cv.fit(X, y)

[multimeric]

@glemaitre, @ogrisel, @thomasjpfan (I'm following this list: https://github.com/scikit-learn/scikit-learn/wiki/Available-reviewers)

[thomasjpfan]

Thank you for the PR.

Before going into a more technical review, we discussion the feature's inclusion in the original issue: #23382 (comment)

[multimeric]

Okay, I've now added an example of this feature. It doesn't rely on the new classes I added, in case it is decided that these classes are not wanted in the repo. If this is decided I will remove them from the PR and keep only the example.

[multimeric]

Here's the rendered version of the example: https://output.circle-artifacts.com/output/job/3166928d-f107-471f-80c5-ea49d5d16d98/artifacts/0/doc/auto_examples/ensemble/plot_oob_cross_validation.html

[multimeric]

Rendered example: https://output.circle-artifacts.com/output/job/5a965e7b-1ff0-487a-a346-a48e07642613/artifacts/0/doc/auto_examples/ensemble/plot_oob_cross_validation.html

[glemaitre]

As mentioned here: #23382 (comment), I would be more in favor of repurposing the already existing example and making it narrative instead of adding a new example. One of the reasons is that we have already too many examples and we should avoid adding a new one when we can improve an existing one (and old one).

[multimeric]

Considering the reduced scope of what you would accept, I don't have any particular interest in finishing this PR, but I will keep the issue #23382 open because I believe it's still an unresolved need in sklearn.