Perform non-blocking socket poll to determine if work is immediately
available, to evenly balance work across multiple processes before
balancing across multiple threads within a single process, which is less
efficient due to Ruby's Global VM Lock.

This makes the loaded accept wait time dynamic, such that less loaded
servers will wait less. That should even out the delay experienced by a
new client connecting to a highly loaded cluster.

This commit also makes the waiting dependent on worker mode with more
than one worker.