It's unlikely that a process will be using restartable atomic sequences
on Alpha, and furthermore it's unlikely that any given context switch
will be returning to one even if the process has them.  So, re-arrange
the RAS processing in cpu_switchto() so that the most likely code paths
are predicted by the branch predictor.  On an EV4-class processor, this
will save ~4-6 cycles on just about every context switch.

(thorpej)