DOC: freeze old whatsnew notes (GH#6856)

[jbrockmendel]

closes #6856

The .. ipython:: blocks in the whatsnew files are re-executed on every doc build against whatever pandas is current. That is why they end up illustrating behavior other than the release they describe, and why they periodically break the warnings-as-errors doc build when a deprecation they demonstrate is enforced — five blocks in v0.18.0 are currently held up by :okwarning: on the inplace deprecation in eval/query alone.

This rewrites those directives as static .. code-block:: pycon sessions across all 45 files that had executable blocks. 799 executable blocks and 49 :okexcept:/:okwarning: markers go to zero.

The conversion was mechanical, but the script that did it is deliberately not part of this PR — it is a one-shot tool that shells out to pixi exec for pandas 0.17.1–2.x under CPython 3.5/3.6, so committing it would leave unmaintainable tooling in scripts/ and 30 tests running in CI for something nobody will run again. It stays available in this PR's history: freeze_whatsnew.py and its tests, at 82d9c06.

To keep the directory frozen going forward, a no-executable-ipython-in-whatsnew pre-commit hook rejects any .. ipython directive under doc/source/whatsnew/. It passes on the tree as of this PR, including the in-development v3.1.0.

Picking up the two requests from the dev call on the previous attempt (GH-64742): >>> prompts rather than In [n]:/Out[n]:, and output produced by the then-current pandas. The second turned out to be more tractable than that discussion concluded — pixi exec reaches back to 0.17.1, not 2.2. Getting era-faithful output needs three things from the release, not just its pandas:

• the source, read back with git show <tag>:<file> — these files have been edited for a decade to keep executing, so v2.0.0 says freq='ME' (a 2.2 spelling) and v0.18.0 passes path_or_buf=/header= to Series.to_csv, neither of which existed. Rendering today's text under 2.0.3 gave 15 tracebacks where 4 were intended.
• the doc header — until 0.24 the whatsnew page did from pandas import *, which is why the released text says Timestamp where today's says pd.Timestamp.
• doc/data, extracted from the tag, so read_csv('data/baseball.csv') reads the era's file.

conda-forge has no pandas below 0.17.1, so the 16 pre-0.17 files are rendered against the current environment instead (--current). Those pages stop breaking the build but are not historically accurate — v0.13.1 now shows RangeIndex and datetime64[us] where 0.13.1 printed Int64Index and no us. Deliberate; happy to revisit if reviewers would rather see a note on those 16.

Seven blocks render as input with no output, because their output was never reproducible: a Timestamp.now() wall clock, three id() calls used as labels, and a Styler whose repr is a memory address. The script detects these by running each file twice from scratch and refusing anything that disagrees with itself, and pins PYTHONHASHSEED so dict ordering under the CPython 3.5/3.6 interpreters those old releases run on is stable. Two independent full runs produce byte-identical output.

Verified: 45/45 pages build (43 clean; the 2 that warn, unknown document: 'v1.2.1' and '/user_guide/io', warn identically on the unmodified files — single-doc mode excluding the referenced pages), pre-commit and mypy clean.

One pre-existing bug is fixed rather than re-rendered: v0.23.0.rst wrote .. ipython without the ::, so docutils treated it as a comment and its sort=False example rendered as nothing. The new hook matches that missing-colons spelling too, so it had to be resolved here; the block is now frozen like its two siblings.

AI disclosure: drafted with Claude Code (claude opus 5 (high)), which wrote the script, ran the freeze, and reviewed the result over three adversarial review rounds.

[mroeschke]

1. I think it would be helpful to have a pre-commit check for the doc/source/whatsnew directory to ensure .. code-block: pycon is the only allowed code-block directive in these files
2. IMO I don't think we'll need to commit freeze_whatsnew.py and the test it's essentially a one time conversion, and we should ensure all whatsnew docs (even the current) are written in a "frozen" state"