Native code due to freethreading

[mikeshardmind]

There's currently threading.RLock use in a few places strictly to have heapq use be safe under freethreading. This isn't a small penalty.

In 3.14+, this should be safe (double check at release, see: python/cpython#135036)

Options here include vendoring the 3.14 copy of the heapq module, implementing this natively with critical sections, dynamically importing versions of the code to only use the lock under freethreading + python < 3.14

[mikeshardmind]

Some light experimentation with this suggests that this may not be worth it short term. The cost of the threading Rlock is actually in the same order of magnitude as the naive minimal native code with critical section use in this case. This probably improves if more of the code is actually native rather than just calling into python, but I haven't checked as that will come up naturally later on if and when making this a native module for other reasons.