Change the way threadDelay# is implemented.

We now use a list of sleeping threads sorted in increasing order by
the time at which they will wake up.  This avoids us having to
traverse the entire queue on each context switch.