gh-117142: ctypes: Migrate global closure freelist to thunk-type state

[neonene]

This introduces PyCThunkType_Type, a metaclass of PyCThunk_Type, to bypass the module check during a finish.

• Issue: Convert _ctypes extension module to multi-phase init #117142

[encukou]

What problem is this solving, and why can't it be solved in a simpler way?

I think the closure freelist can stay process-global. The free-threaded build will need some locking (not just for multiple interpreters), but I think closure alloc/dealloc happens rarely enough that a global mutex should be enough. (This is unlike most of Python's freelists, which exist because allocation/deallocation is too common.)

We don't free the allocated pages at interpreter shutdown. I don't think that's a big concern, though -- while we don't munmap the pages, we do recycle them.

Is there something I'm missing?

[neonene]

I'm trying to reduce global variables (globals-to-fix.tsv) as much as possible first, then asking for a decision. If you think that the closure (de)allocation is not frequent, malloc_closure.c (freelist) can be removed in the future? Leaving a freelist-less experiment. The CI appears to be OK.

[encukou]

It would be nice if libffi (ffi_closure_alloc/_free) always did all this for us (on all platforms) so we could drop this code. Not sure what's the status there; I don't know that much about libffi.