[backport 1.11] Set array size only when safe to do so

[MilesCranmer]

I looked through the _growend! function from the stacktrace here and noticed that the .size of an array might be getting set too early; before the memory was actually allocated. Maybe there are very specific GC conditions (juliacall + macOS) where this segfaults things.

Possible fix for #58475 #58171. Edit: nope, this wasn't the fix. But seems like a good thing to have anyways so I'll submit it regardless

[bvdmitri]

Do you mean potential race-conditions with GC when saying that the current behaviour sets size "too early"? If this is the case, the proposed behavior is still prone to race conditions, but it will be "too late" instead because the memory has been allocated but the size has not been updated yet. Those two fields really should be set together atomically, the actual order should not be relevant

[MilesCranmer]

I agree. BUT since they aren't atomic, this order is the least bad.

If the array shrinks, then you modify the length immediately. If the array grows, then you modify the length at the end.

This lowers the chance of (but does not completely remove) corruption from thread race.

I don't feel strongly though. Feel free to take or reject

[mbauman]

Let's tackle #58848 first and then come back to backport it after that is merged. You should be able to re-open this one at that point... if a manual backport is even necessary (if it applies cleanly then it can be automatically hoovered up into a batch of backports in one go).