Support genuine free-threaded mode

This complex task covers many C functions for MultiDict (and family) methods and slots.

Some of them are trivial and could be solved by wrapping the call in a critical section, for example, ' repr '.

But applying critical sections to all methods kills performance.

The pragmatic approach should combine critical sections with atomic operations, as CPython does for `dict`.

I expect a bunch of PRs for this issue. Let's eat the elephant piece by piece.