Extend Windows Process completion key's lifetime

[HertzDevil]

Alternative to #14995. Fixes #15028.

There is probably a better data structure than Set here...

[ysbaddaden]

Accesses to CompletionKey.all should be protected by a Thread::Mutex when :preview_mt if set.

[ysbaddaden]

Thank you 🙇

[straight-shoota]

I suppose this might be fine for getting the bug fixed, but it seems far from ideal.

We probably shouldn't need this for all CompletionKey instances if it's primarily an issue with Process.
And I'm particularly concerned about @@all becoming a memory leak because it never scales down.

[HertzDevil]

That would be a use case for #15494 (so would the Deque used for the fiber stack pool, although that consumes far less memory than the stacks it holds).

I was contemplating some kind of linked list but don't think O(n) deletion is worth it.

[straight-shoota]

Deletion in a doubly linked list would be O(1) if we have a pointer to the list node. It could be just the CompletionKey instance itself.

[ysbaddaden]

Indeed, it's something for Thread::LinkedList with intrusive nodes.

[ysbaddaden]

Thank you 🙇