TST: Use importlib for importing in pytest

[QuLogic]

PR summary

This fixes the rewriting so that we get detailed variable information again, which was broken with the move to Meson (which does weird import hacking for editable installs.)

PR checklist

• [n/a] "closes #0000" is in the body of the PR description to link the related issue
• new and changed code is tested
• [n/a] Plotting related features are demonstrated in an example
• [n/a] New Features and API Changes are noted with a directive and release note
• [n/a] Documentation complies with general and docstring guidelines

[QuLogic]

Interestingly, this seems to be fine on the Minimum Versions CI, but everywhere else doesn't like it. That build uses pytest 7.0.0, and locally I have 7.2.1. The other jobs appear to be 7.4.3. I will have to investigate whether that might be the cause of the problem.

[QuLogic]

And on Windows, I have pytest 7.4.0 which works.

[QuLogic]

And updating my Windows machine to 7.4.3 also works, so it seems to be something else.

[tacaswell]

If I launch the tests as

pytest -n 20

it works

If I launch the tests as

python -mpytest -n 20

it fails locally

python -I -mpytest -n 20

works.

[QuLogic]

Ah right; I think what we'd want is python -P? But it wasn't available in all Python versions we support. I'm also confused why the Minimum Versions was okay with that since it doesn't run things any differently.

But to continue, do we want to:

1. Change only CI to use importlib and direct pytest?
2. Stop recommending python -m pytest over pytest (note their default documented invocation is with pytest, not python -m pytest)?
3. Change the pytest options for everyone?

[tacaswell]

I think we want to:

• change the options for everyone (not getting the diffs is annoying)
• change how we invoke the tests on CI and in our docs to just pytest
• add a note about python -P / python -I for running with python -m

There was a big push to move everything to python -m ... but that seems to have fizzled out. The main argument for it was to avoid the risk of picking up cli tools from the underlying shell when in a conda/virtual environment. However, this seems to have been instead handled by a combination of "don't mess up", aggressively always installing pip (which had the highest risk of causing chaos!), and better tooling around managing environments.

[QuLogic]

OK, now we can see that that leak test if failing as:

>       assert growth <= acceptable_memory_leakage
E       assert 3014656 <= 2000000

It appears that Cygwin broke a while ago, and I've opened pybind/pybind11#4999 for that.