Make ffi_init_once thread-safe in free threaded build

[colesbury]

Lock the FFI object when initializing the init_once_cache dictionary and use PyDict_GetItemRef in places where the dictionary's value may be concurrently modified.

See #126

[colesbury]

Hi @nitzmahone - I'm still working through the CFFI free-threading issues, but I figured it'd be helpful to get some feedback on a small change.

On that note:

• I added a "polyfill" for PyDict_GetItemRef, but that sort of thing is nicely provided. by https://github.com/python/pythoncapi-compat/. Would you be interested in using that?
• It's probably cleaner and simpler to make init_once_cache have the same lifetime as its FFIObject instead of lazily initializing it. I went with the smaller code change for now, but let me know what you think.

To find thread-safety issues, in addition to reading through the code, I'm using https://github.com/Quansight-Labs/pytest-run-parallel, which I've found to be very effective.