Add support for using ephemeral ports

[muggenhor]

Ephemeral ports are used when binding to port '0'. The operating system
then automatically selects a (non-zero) port for you that's unused at
the time. This is great to prevent conflicts.

This commit extends the wire server such that it reports the actual port
number we've bound to, even when using ephemeral ports. This is useful
to prevent port conflicts on CI systems running many tests in parallel.

[muggenhor]

Without this it's possible for the step definition executable to still be running when cucumber is finished. Keeping the port in use, if timing is working against us just enough then this can result in an inability to open a new listening socket on that port for the BOOST test case below.

Note that a similar race condition still exists here: the step definition executable will not necessarily have opened the TCP socket for listening by the time that cucumber attempts to connect to it. An approach that I'm using in our CI system is to run with --verbose and wait for the Listening on ... message before starting cucumber. That's more difficult to script in a shell script than the Python that I'm using though. (I'm using Python instead of Ruby because most of our CI system is either shell scripts or Python).

[paoloambrosio]

Thanks for the contribution! Apart from the comment on SocketServer::listenPort() I think it can be merged.

[paoloambrosio]

I don't think any part of the software can handle a failure here (return value of 0). Isn't it better to use the other signature that throws boost::system::system_error to make it crash if for some reason the port number cannot be retrieved?

If you agree, remember to change the documentation for the interface as well :-)

P.S. Position of { not consistent with the rest of the code

[muggenhor]

@paoloambrosio: ack, it's in with the new push

[paoloambrosio]

One of the Travis builds failed but there is nothing wrong with this PR. The build should go back to green when the next PR is merged.