MNT Add nightly workflow that tests the latest released version - #1809

Description

Fixes #1188.

This adds a test-latest-release.yml workflow and calls it from nightly.yml. It looks up the latest Fairlearn version on PyPI, checks out the matching v<version> tag, installs it with the current versions of the dependencies and runs the same pytest command that the release itself used in test-all-deps.yml (without coverage). This way a dependency release that breaks the version users actually install shows up in the nightly run, not only through bug reports.

Following the discussion in the issue, it installs from the checked-out release tag (pip install .) rather than from PyPI, so that the tests and the package come from the same commit. A step checks that the installed fairlearn.__version__ matches the PyPI version. The workflow also has workflow_dispatch, so it can be started by hand for a quick check after a dependency release.

I could not run the workflow on GitHub Actions from my fork (no runs were created for it), so I ran the same steps on an Ubuntu 24.04 machine with Python 3.12 instead. Against today's dependencies (pandas 3.1.1, numpy 2.5.4, scikit-learn 1.9.1, scipy 1.18.1), the v0.15.0 tests give 2408 failed, 4523 passed. The failures I looked at come from pandas.errors.Pandas4Warning: Setting a value on a Series via .at with a key that does not exist in the index is deprecated, which is turned into an error by filterwarnings = ["error"]. This is already fixed on main by #1801 but not released yet, so the new job would currently be red. That is the kind of signal this job is meant to give.

Tests

• no new tests required
• new tests added
• existing tests adjusted

Documentation

• no documentation changes needed
• user guide added or updated
• API docs added or updated
• example notebook added or updated