[BUG?] Investigate shared ABI and module state

[da-woods]

Describe your issue

1. Do we need to add CYTHON_USE_MODULE_STATE to the CYTHON_ABI suffix? I think it might be OK without but there's the potential to go wrong.
2. Is there a lifetime issue? In principle modules can now be unloaded (although in practice I don't think our reference counting is accurate enough for this to happen...). If we store CyFunction in our shared ABI based on loading it from module a, and then unload module a does this cause a problem? I think it might.
3. __Pyx_Coroutine_get_frame will always get the globals from whichever module added it to the shared ABI.

I think this is all mostly a theoretical problem, but it's all a bit dubious.

[da-woods]

For point 2 - I think we need to disable internal type-checks with CYTHON_USE_MODULE_STATE. i.e. it's fine for other utility code to check if something is a CyFunction, but it's not fine for CyFunctions to check if something is a CyFunction (although for CyFunction specifically there might be a way of doing it).

I think that mainly affects Coroutines and related types. I think these checks are only an optimization though so they're OK to drop.

[da-woods]

I'm also not convinced __Pyx_Coroutine_Check and __Pyx_Coroutine_CheckExact are used correctly in all cases even without module state. They're mostly just used for optimization (where it doesn't really matter if it misses) but there's a few cases where it affects behaviour. Can't see an obvious way around that though.