Apply pending refcnt updates before release

[adamreichold]

Closes #4105

[davidhewitt]

I think based on #4105 (comment), this fix probably isn't enough. My test case shown in that comment still crashes.

[adamreichold]

I think based on #4105 (comment), this fix probably isn't enough. My test case shown in that comment still crashes.

Do we still want this for a 0.21.3 bugfix release? It does not fix everything but it should be closer to the behaviour with delayed decrements based on the GIL pool.

[davidhewitt]

Given that it looks like we can't land a full fix in a patch, sure, let's do this and then I'll see about getting a patch shipped soon 👍

[adamreichold]

I have reconsidered this and I think we should not do this at all: It does not seem to solve the OP's problem, can significantly increase our call overhead, we have to remember to remove this after solving this issue for good and wrapping things in allow_threads should provide a workaround for existing code.

After #4095 is finished, it should not be a correctness issue any more when/where we call update_counts and we can revisit whether we want to call it each and every time or only if a certain amount of updates is pending etc.