API Revamp estimator tags

[adrinjalali]

Closes #22606
Closes #20804

This PR revamps estimator tags, puts them in dataclasses, and is based on #22606

High level changes from this PR:

(from #22606):

• replace MRO mechanism with inheritance
• remove _get_tags and _more_tags and introduce __sklearn_tags__

this PR:

• dataclasses are introduced to store the tags, and they're scoped in a few dataclasses. This helps users with auto-complete as well.
• stateless is removed and now we only use requires_fit. The two were redundant.
• only_binary is now replaced with multi_class
• multioutput_only is removed and now we have multi_output and single_output
• get_tags, default_tags, and Tags are put into public API

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: 88af896. Link to the linter CI: here}

[glemaitre]

A bunch of changes to be consistent for the documentation style mainly.

[adrinjalali]

It's a green CI @glemaitre

[glemaitre]

So LGTM.

[larsoner]

I had a quick look at the docs and maybe I missed it, but as a consumer of sklearn that subclasses BaseEstimator I'm not sure how to adapt my code. Fore example, I currently have classes that have stuff like:

    def _more_tags(self):
        return {"no_validation": True}

How do I adapt these classes in a way that is backward compatible with previous sklearn versions? I can't just leave _more_tags in there to be picked up by older versions of sklearn because then I fail validation:

mne/decoding/tests/test_search_light.py:373: in test_sklearn_compliance
    check(est)
../virtualenvs/base/lib/python3.12/site-packages/sklearn/utils/estimator_checks.py:3893: in check_estimator_tags_renamed
    assert not hasattr(estimator_orig, "_more_tags"), (
E   AssertionError: ('_more_tags() was removed in 1.6. Please use __sklearn_tags__ instead.',)

But I can't remove it because then it won't be backward compatible.

One solution would be to add an opt-in to have the validator ignore this attribute being present, or maybe change it to ensure that if _more_tags is present then __sklearn_tags__ is also present.

Or is there a simpler way for me to adjust my code?

[adrinjalali]

@larsoner you can leave _more_tags there with an @available_if decorator, and the check would be the scikit-learn version.

Something like this:

import numpy as np
import sklearn
from packaging import version
from sklearn.base import BaseEstimator
from sklearn.utils.estimator_checks import parametrize_with_checks
from sklearn.utils.metaestimators import available_if

from sklearn.utils.validation import check_is_fitted, validate_data

def check_version(estimator):
    return version.parse(sklearn.__version__) < version.parse("1.6.dev")

class MyEstimator(BaseEstimator):
    @available_if(check_version)
    def _more_tags(self):
        return {"_skip_test": False}
    
    def fit(self, X, y=None):
        validate_data(self, X, y)
        return self
    
    def predict(self, X):
        check_is_fitted(self)
        validate_data(self, X, reset=False)
        return np.zeros(X.shape[0], dtype=int)
    
@parametrize_with_checks([MyEstimator()])
def test_my_estimator(estimator, check):
    check(estimator)