NB: Changes to parameter defaults & units included in this commit!

I tried to apply some queuing theory to this problem and as much as
I admire Agner Krarup Erlang, his math isn't much help when you don't
know any of the relevant metrics for your queue.

Instead I took a much simpler approach:  "If we fail to get a thread,
we probably need more threads", and have rewritten the herder to to
react faster and more reliably to such events.

I went over the parameters for thread-pools and normalized timeouts
to seconds rather than milliseconds (beware!) and polished descriptions
etc.