The test was not waiting for all the worker-created sockets
to be listening before calling cluster.disconnect().
As a result, the channels with the workers could get closed
before all the socket handles had been passed to them, leading
to various errors.

When a child process closes the channel to the parent, we should
call _disconnect(), which raises the 'disconnect' event immediately,
rather than disconnect(), which can potentially be deferred if
there is a handle queue on the channel. A deferred disconnect would
never be processed because the other side is gone and the channel
will never be drained.

…cha-testing