chore: avoid importing TYPE_CHECKING

[mayeut]

Using TYPE_CHECKING=False rather than importing it allows not to import typing altogether which can help with import time.
This PR forbids the use of typing.TYPE_CHECKING to ensure typing will not be imported if unneeded.

[notatallshaw]

This is pointless code churn

[notatallshaw]

Provide a real world use case that uses packaging and markedly improves due to saving type importing.

I am -1 on this without a real world benefit, while this works in practice isn't not part of the typing standard and adds a lot of churn across the code base.

[mayeut]

you're right in the end it always typing always ends up being loaded. Let's wait for PEP781 to properly land in a few years. sorry for the noise.

[henryiii]

For context, this is probably inspired by pypa/cibuildwheel#2866, which itself was inspired by henryiii/cibuildwheel#17 (from @hugovk). For cibuildwheel, when combined with lazy imports, it saves about 30% more time, and simplifies the lazy imports a bit (you can't lazy import TYPE_CHECKING and then use it, it reifies immediately). Not sure it's as effective here, though.

[mayeut]

Not sure it's as effective here, though.

It's only effective for the errors & direct_url modules which is likely useless on its own . I should have checked that before hand.
There's a bunch of types exposed publicly in other modules and it would probably be deemed a regression to only expose them under the guard in order for this to be potentially effective in real-world applications.
typing import is roughly 1-2 ms (when all its deps are already imported) on Apple M1 / Python 3.15.