server: handle startup errors more nicely

[oliver-sanders]

If an error occurs during the startup of the workflow network server we get a long nasty traceback irrespective of the error type.

The reason for this is that exceptions raised inside of threads cannot be caught from the parent thread so Cylc eventually falls over when the timeout on the "barrier" is hit.

There is one exception which is relatively likely to occur on startup:

zmq.error.ZMQError: Address already in use

This can happen if all of the ports in the specified range are occupied, OR, I guess potentially if multiple workflows/processes try to claim the same port simultaneously, dunno. To replicate this add the following to your global.cylc:

[scheduler]    
    [[run hosts]]    
        ports = 43042 .. 43042

Ideally we would, somehow catch this exception in the parent thread and pass it through our standard error handling which aught to niceify the error a bit making it more obvious what the issue is.

There's some info on catching exceptions in threads here:

https://stackoverflow.com/questions/2829329/catch-a-threads-exception-in-the-caller-thread

Looks like the ThreadPoolExecutor passes the exceptions around nicely - https://stackoverflow.com/a/12808634

tldr; we don't want the exception TB to arise from the barrier.wait if possible.

Pull requests welcome!

[oliver-sanders]

(note same situation on master as in #4274)