As discussed at the developer meeting.

Close #27662, close #28956.

Here is a summary of the motivations:

There is very little scikit-learn + PyPy usage

• No PyPy + numpy usage were reported in this HN question by @mattip.
• No user said "yes I used scikit-learn + PyPy and I care about this" in Olivier's social media posts, see PyPy tests timeouts / memory usage investigation #27662 (comment).
• the only way to install scikit-learn with PyPy without building from source is conda-forge. There is no scipy PyPy wheel and there is no scikit-learn PyPy wheel.
• last available scikit-learn version on conda-forge is 1.0.2 and noone has ever complained about this in the last two years and a half.
• Download stats on conda-forge are quite low, see https://anaconda.org/conda-forge/scikit-learn/files?version=1.0.2. The PyPy scikit-learn package has py37hca48384_0 in its filename (mamba create -n pypy scikit-learn tells you this) and has been downloaded ~2000 times to be compared to the other downloads that amounts to ~3 million downloads.

Maintenance has a cost

• historically PyPy tests have been timing out regularly likely because of memory issues and investigation is extremely time-consuming, see PyPy tests timeouts / memory usage investigation #27662
• even now that we run only the common tests, the PyPy lock-file update fails after a dependency update for some reason 🔒 🤖 CI Update lock files for pypy CI build(s) 🔒 🤖 #28956.
• conda-forge package is using Python 3.7 and has not been updated in roughly two years and a half. There are some issues that nobody has had time to investigate, see e.g. Rebuild for PyPy3.9 conda-forge/scikit-learn-feedstock#248 (comment).

Changes in this PR

• Update of the FAQ to say that PyPy is not officially supported
• Remove the CI build and lock-files
• Update of the scikit-learn code that had some PyPy-specific code. These changes could be reverted if we want to say: "PyPy is not tested but simple fix PRs may be accepted"
• update of the tests in particular the PyPy-specific markers to skip some tests when run under PyPy

As things stand in this PR, PyPy is mentioned only in the FAQ, some old changelog, setup.py (no need to tweak this since setup.py will go away soonish), and because sklearn.utils.IS_PYPY is deprecated and scheduled to be removed in scikit-learn 1.7:

❯ git grep -i pypy
doc/faq.rst:Do you support PyPy?
doc/faq.rst:scikit-learn with `PyPy <https://pypy.org/>`_ (an alternative Python
doc/whats_new/v0.20.rst:- |Feature| Add almost complete PyPy 3 support. Known unsupported
doc/whats_new/v0.20.rst:  :class:`feature_extraction.text.HashingVectorizer`. For running on PyPy,
doc/whats_new/v0.20.rst:  PyPy3-v5.10+, Numpy 1.14.0+, and scipy 1.1.0+ are required.
doc/whats_new/v1.1.rst:- |Feature| :class:`feature_extraction.FeatureHasher` now supports PyPy.
doc/whats_new/v1.2.rst:  `nu=0.5` for PyPy (and possibly other non CPython interpreters). :pr:`24245`
doc/whats_new/v1.4.rst:- |Fix| fixes a memory leak seen in PyPy for estimators using the Cython loss functions.
doc/whats_new/v1.5.rst:- |API| :data:`utils.IS_PYPY` is deprecated and will be removed in version 1.7.
setup.py:            "compile_for_pypy": False,
setup.py:    is_pypy = platform.python_implementation() == "PyPy"
setup.py:            if is_pypy and not extension.get("compile_for_pypy", True):
sklearn/externals/_arff.py:- Fully compatible with Python 2.7+, Python 3.5+, pypy and pypy3;
sklearn/utils/__init__.py:    if name == "IS_PYPY":
sklearn/utils/__init__.py:            "IS_PYPY is deprecated and will be removed in 1.7.",
sklearn/utils/__init__.py:        return platform.python_implementation() == "PyPy"
sklearn/utils/tests/test_utils.py:def test_is_pypy_deprecated():
sklearn/utils/tests/test_utils.py:    with pytest.warns(FutureWarning, match="IS_PYPY is deprecated"):
sklearn/utils/tests/test_utils.py:        from sklearn.utils import IS_PYPY  # noqa