ci: Test with the 3 latest major versions of scikit-learn

[thomass-dev]

Refactor skore.yml workflow:

• launch tests with the 3 latest major versions of scikit-learn on os: "ubuntu-latest" and python: "3.12",
• speed-up tests by installing dependencies in venvand caching it.

And make _add_scorers compatible with sklearn ==1.4.

[augustebaum]

Not sure skore is critical to the point of making our CI run 4 times slower

[github-actions]

pytest coverage report

File	Stmts	Miss	Cover	Missing
src/skore
__init__.py	12	0	100%
__main__.py	8	1	1	80%
exceptions.py	3	0	100%
src/skore/cli
__init__.py	5	0	100%
cli.py	33	0	100%
color_format.py	43	3	2	90%
launch_dashboard.py	26	15	0	39%
quickstart_command.py	14	0	100%
src/skore/item
__init__.py	21	0	100%
cross_validation_item.py	137	10	2	93%
item.py	41	13	0	68%
item_repository.py	42	2	1	93%
media_item.py	70	4	1	94%
numpy_array_item.py	25	1	1	93%
pandas_dataframe_item.py	34	1	1	95%
pandas_series_item.py	34	1	1	95%
polars_dataframe_item.py	32	1	1	94%
polars_series_item.py	27	1	1	94%
primitive_item.py	27	2	1	92%
sklearn_base_estimator_item.py	33	1	1	95%
skrub_table_report_item.py	10	1	1	86%
src/skore/persistence
__init__.py	0	0	100%
abstract_storage.py	22	1	0	95%
disk_cache_storage.py	33	1	1	95%
in_memory_storage.py	20	0	100%
src/skore/project
__init__.py	3	0	100%
create.py	52	8	0	88%
load.py	23	3	0	89%
open.py	14	0	100%
project.py	64	4	4	91%
src/skore/sklearn
__init__.py	3	0	100%
find_ml_task.py	35	3	2	89%
types.py	2	0	100%
src/skore/sklearn/cross_validation
__init__.py	2	0	100%
cross_validation_helpers.py	47	4	1	90%
cross_validation_reporter.py	35	1	1	95%
src/skore/sklearn/cross_validation/plots
__init__.py	0	0	100%
compare_scores_plot.py	29	1	2	92%
timing_plot.py	29	1	1	94%
src/skore/sklearn/train_test_split
__init__.py	0	0	100%
train_test_split.py	34	2	1	94%
src/skore/sklearn/train_test_split/warning
__init__.py	8	0	100%
high_class_imbalance_too_few_examples_warning.py	17	3	2	78%
high_class_imbalance_warning.py	18	2	1	88%
random_state_unset_warning.py	11	1	1	87%
shuffle_true_warning.py	9	0	1	91%
stratify_is_set_warning.py	11	1	1	87%
time_based_column_warning.py	22	1	2	89%
train_test_split_warning.py	5	1	0	80%
src/skore/ui
__init__.py	0	0	100%
app.py	25	5	2	71%
dependencies.py	7	1	0	86%
project_routes.py	50	0	100%
src/skore/utils
__init__.py	0	0	100%
_logger.py	21	4	0	84%
_show_versions.py	31	0	100%
src/skore/view
__init__.py	0	0	100%
view.py	5	0	100%
view_repository.py	16	2	1	83%
TOTAL	1380	107	91%

Tests	Skipped	Failures	Errors	Time
259	0 💤	0 ❌	0 🔥	34.149s ⏱️

[thomass-dev]

Not sure skore is critical to the point of making our CI run 4 times slower

Older versions of scikit-learn are only tested on ubuntu-latest python 3.12.

[glemaitre]

So as a workaround for the older scikit-learn function, I think that you can:

• add a parameter estimator to _add_scorers
• pass self.estimator to the call to _add_scorers in the CrossValidationReporter.

Basically, in the case that scoring is None, we would default to self.estimator.score. However, in practice it will never happen because we are creating a list of scorer to scoring is never None. So it just a workaround.

I think it is worth a small comment to know why we pass the estimator around and actually don't expect to use it.

[glemaitre]

I open #1062 that should solve the issue.

[thomass-dev]

Thanks @glemaitre for your fix, all good. It's ready to merge @rouk1 .

[rouk1]

Nice work, the planet says thx for caching venvs 💚