Previously, the server would attempt taking port/socket in worker
process, and if it was taken, fail, which made the master process
create a new worker. This led to really high CPU usage if the
streaming API was started when the port or socket were not
available.

Now, before clustering (forking) into worker processes, a test
server is created and then removed to check if it can be done.