Just a long shot for a failure observed on gorakhargosh#998. My hypothesis is that
when we stop ProcessWatcher before we restart the process manually, we
don't yield to it and immediately kill the process. Next, when the
ProcessWatcher thread is woken up, we have to conditions ready - the
popen_obj and stopped_event, see the corresponding code, ``` while
True: if self.popen_obj.poll() is not None: break if
self.stopped_event.wait(timeout=0.1): return ```

And desipte that `stopped_event` is set, we first check for
`popen_obj` and trigger the process restart.

We can also make the ProcessWatcher logic more robust, by checking if
we are stopped before calling the termination callback, e.g.,

```
        try:
            if not self.stopped_event.is_set():
                self.process_termination_callback()
        except Exception:
            logger.exception("Error calling process termination callback")
```

I am not 100% sure about that, as I don't really know what semantics
is expected from ProcessWatcher by other users. But at least the
AutoRestarter expects this semantics - i.e., a watcher shall not call
any events after it was stopped.