Add locking between replicaGCQueue and multiraft.state.createGroup.

[bdarnell]

This partially addresses the race seen in #2815. A similar race still
occurs but much less frequently.

[bdarnell]

I'm not sure how to test this. I've been testing it locally by increasing the iteration count in TestRaftRemoveRace, although checking in that change would noticeably increase total test runtime. I don't see a good way to trigger the race in a more direct/controlled way.

[tamird]

remove?

[bdarnell]

Done.

[tamird]

What is the similar race? can you document it?

[bdarnell]

I'm still working on identifying the similar race. All I know so far is that it produces the same error message as #2815 and it takes over a minute for the test to reproduce it.

[bdarnell]

One "similar race" is that I forgot to return nil after the "aborting deletion" log line. But even with that fixed I'm seeing other rare failures. I suspect that what may be happening is that the node is sometimes falling far enough behind that it is learning about multiple iterations of the add/remove loop at once.

[tamird]

LGTM