When using pytest, I've noticed that it's harder to diagnose test collection errors than test failures (most recently in #2869 & #2870). Test collection errors don't result in good tracebacks, at least with the pytest settings from pyproject.toml. Test collection errors also prevent pytest from running the full test suite by default, which means we often get incomplete information.

In this PR, I'm planning to add some tips to the testing guide on how to avoid getting test collection errors. In particular, if a test file contains code that has a chance of failing, that code should be put into either a fixture or an actual test. Pretty much I want to avoid situations like...

# test_file.py

x = 0 / 0   # causes a ZeroDivisionError in the test collection phase

def test_something():
    assert True

I may also look into ways to improve the reporting of test collection errors.

Just wanted to start this PR now as a placeholder since it might take me a month or three to get to it.