Two changes:

  - newly created or unblocked threads go to the back of the run
    queue.  This is fairer because it avoids possible starvation.
    Fixes problem with conc004 test going much slower with optimisation.

  - The context switch flag is only reset after a yield.  Previously,
    it was being reset before we ran a thread, which meant that a tick
    that arrived during GC would be lost, and a thread doing lots of
    GC could hog the CPU.  Now, GC is effectively counted as part of
    the last running thread's timeslice.

MERGE TO STABLE.