HANG waiting for thread_initexit_lock from self-interp of new managed thread during attach

Another attach race hit with the test from #2601 (closed).

The loop that takes over new threads lets taken-over threads start running, and they can go create child threads. If a managed child has not yet initialized and is waiting for the thread_initexit_lock at the top of dynamo_thread_init, the attach code will send it a takeover signal and we'll interpret our own code.

I'll spare the gory debugging details: it took some time to figure out what was going on, but this happens on a regular basis, at least once every 20 runs of the new test.