GH-46516: [CI][Python] Force Cython>3.1.1 for docs builds

[raulcd]

Rationale for this change

Due to a change of behaviour on Cython for 3.1 the docstrings generated fail when being tested with numpydoc. The issue was fixed on Cython==3.1.2.

• [BUG] Docstring pre-clean breaks numpydoc integration due to removal of \n\n cython/cython#6904

What changes are included in this PR?

• Pin Cython>3.1.1 for the docs requirements (both conda and pip)
• Fix doctest failure that was introduced in the interim when the CI job had been failing.

Are these changes tested?

Via CI.

Are there any user-facing changes?

No

• GitHub Issue: [Archery][CI][Doc] archery numpydoc fails with latest Cython 3.1.1 #46516

[github-actions]

⚠️ GitHub issue #46516 has been automatically assigned in GitHub to PR creator.

[pitrou]

Will that work properly if Cython was installed using conda?
Also, is this necessary at all?

[raulcd]

Will that work properly if Cython was installed using conda?

It does fix the existing failures which install Cython through conda, so the answer seems to be yes (without this it still fails) because the cached docker image installed Cython 3.1.1

Also, is this necessary at all?

I don't think we should merge that specific change, I was validating the existing failing CI jobs.
I think the bests course of action would be to force rebuild the underlying docker image so it is not caches and it installs Cython 3.1.2. I am asking on Zulip if someone has access to docker.io/apache/arrow-dev as rebuilding the image would fix the issue

[pitrou]

LGTM, just one suggestion

[kou]

+1

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 4 benchmarking runs that have been run so far on merge-commit d2a8b78.

There were 70 benchmark results with an error:

• Commit Run on arm64-t4g-2xlarge-linux at 2025-06-10 22:40:34Z
  • tpch (R) with engine=arrow, format=parquet, language=R, memory_map=False, query_id=TPCH-07, scale_factor=1
  • tpch (R) with engine=arrow, format=native, language=R, memory_map=False, query_id=TPCH-09, scale_factor=1
• and 68 more (see the report linked below)

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about 49 possible false positives for unstable benchmarks that are known to sometimes produce them.