Confusing Nanny on_exit callback structure

The Nanny is managing a subprocess in which a Worker is started. If that process exits a cascade of on_exit callbacks are triggered.

The order in which things happen is

- `AsyncProcess._on_exit`
This one is not doing a lot. It sets an event s.t. the process is joinable. Then it triggers another on_exit
- `WorkerProcess._on_exit` is just calling `WorkerProcess.mark_stopped`
- `WorkerProcess.mark_stopped` is resetting some state in `WorkerProcess` and is calling another on_exit
- `Nanny._on_worker_exit_sync` is scheduling a coroutine on the loop which is the next on_exit
- `Nanny._on_worker_exit` is unregistering the worker from the scheduler and if need be restarts the worker process

This chain of events is not only confusing but also subject to race conditions. Particularly that the final, most relevant on_exit callback is scheduled with a `loop.call_soon` allows for various race conditions.

These race conditions are currently not a direct issue. Most race conditions are actually buffered by various idempotent implementations of close/start but once we touch this structure, this is getting a bit shaky.

I debugged this during the investigation of https://github.com/dask/distributed/issues/7312 but this chain is not directly causing the issue. This issue is mostly to document the situation.