Move PyErr_NormalizeException() up a few lines

[rwgk]

Description

The only production code change is to move PyErr_NormalizeException() up a few lines in detail::error_string().

The 1st commit (e564142) purely adds test coverage, without making changes to the production code. This effectively documents the current error_already_set::what() behavior.

The 2nd commit (254bfeb) moves PyErr_NormalizeException() up a few lines in detail::error_string() and adjusts the new tests accordingly, effectively documenting the behavior change, which can be summarized as:

• It fixes the "happy" tests.
• The behavior for the two corner cases has more integrity but is still not really great. This was already addressed under PR error_already_set::what() is now constructed lazily #1895.

Note that all existing tests pass as-is! Only the newly added tests need adjustments.

Background: The production code change was adopted from PR #3964. The new tests are based on similar code under PR #1895. PR #1895 has several behavior changes. This PR handles one behavior change in isolation, for easier review and a clearer development history.

Suggested changelog entry:

``error_already_set::what()`` now handles non-normalized exceptions correctly.

[rwgk]

Looks like a GHA infrastructure bug popped up in the few minutes between the last two CI runs.

I'll try again a little later, but the only thing that changed was the extra std::move(), which seems to work for all non-Windows CI jobs.

Run jwlawson/actions-setup-cmake@v1.12
with:
github-api-token: ***
use-3[2](https://github.com/pybind/pybind11/runs/6599035794?check_suite_focus=true#step:6:2)bit: false
env:
PIP_ONLY_BINARY: numpy
FORCE_COLOR: [3](https://github.com/pybind/pybind11/runs/6599035794?check_suite_focus=true#step:6:3)
PYTEST_TIMEOUT: 300
pythonLocation: C:\hostedtoolcache\windows\Python\3.9.12\x6[4](https://github.com/pybind/pybind11/runs/6599035794?check_suite_focus=true#step:6:4)
Error: Could not find win32 asset for cmake version 3.23.2

[rwgk]

Close-opening to re-trigger CI, hoping the GHA infrastructure failure was resolved in the meantime.

[rwgk]

@Skylion007 thanks for your suggestions, they are applied now. I also answered the question about the PyPy segfault (yes, even without the production code change). The CI is green again. Could you please take another look?

[Skylion007]

@rwgk Is this still a draft PR?

[rwgk]

@rwgk Is this still a draft PR?

No sorry, I keep forgetting to click the Ready for review button.

[Skylion007]

LGTM, merge it.

[rwgk]

Thanks, done! I'll work on updating #1895 (I assume there will be merge conflicts).