gh-158406: Fix run_until_complete making the next run stop early

[tjkuson]

Each run_until_complete call registers its own done callback that only stops the loop whilst that call is running. It no longer stops a run after the call returns.

I'll flag that AnyIO currently finds its main task by looking for asyncio.base_events._run_until_complete_cb in task callbacks. On this branch, it no longer finds it, so anyio.to_thread starts a new thread per task and anyio.to_process jobs can fail. I think matching on the callback's __module__ in AnyIO would fix it.

I also tried an alternative that records the current future on the loop instead. It keeps AnyIO working, but adds loop state, hangs in two edge cases, and doesn't fix a later run_forever().

A neater solution would inline _run_until_complete_cb into the closure and drop its gh-66619 early return, which is now redundant, but I kept it because removing it make AnyIO fail at import.

Closes gh-158406.