A failed log append in a worker leaves records and writes pending forever

[SimonHeybrock]

If appending an event to the log raises in a worker (for example a full disk), the executor swallows the exception and the backend's state stays as before the append. A record whose Finished failed to append stays pending; with persist (#52), a write whose Written failed to append stays pending too. Waiters then hang: client.output of another client, client.wait, and close(), which waits for pending writes.

The log is the backend's only storage, so a failed append means the backend no longer knows its own state. Options:

• the backend stops: every waiter and every later call raises with the append's error;
• the worker retries the append, then stops as above.

Found in the review of #52.