The unit test WorkerSleepingTest.test_stop_current_task_go_sleep fails
on 3.5.3.

This allows for some slight network lag when the job storage resides
on a networked file system.

When the worker receives the "shutdown" state it'll cleanly shut down, and
eventually go to "offline" state.

This is only supported on POSIX platforms that have those signals, i.e.
not on Windows.

In this change the state of the worker is more explicitly stored in
FlamencoWorker.state.

The Manager can now dictate that the Worker moves to an 'asleep' state.

When the "may I run" system says "no", we shouldn't push a "canceled"
status to the Manager, as this could overwrite another valid status that's
there for some other reason.

This refusal can be because another worker is working on the task, but
can also be for other reasons (like the task no longer being runnable).

This way it can be called from multiple coroutines without getting race
conditions.

It'll be called in non-shutdown situations too.

Also changed from getting the local IP address to just using 0.0.0.0, which
seems to work too.

Apparently it's not supported on Windows when using IPv6.

This would influence all socket timeouts on all network connections we do,
which is a bad idea.