This one fixes up the IPI code to do something more sensible.  Sorry,
was just too ugly to leave it alone ...  but I did keep it seperated out
;-) Though this is not an equivalent transform it will only affect
NUMA-Q & summit - same op twice because some twit just split it out in
the last patch for both NUMA-Q & Summit.

Because clustered apic logical mode can't do arbitrary broadcasts of
addressing (it's not just a bitmap), I have to do send IPI instructions
as a sequence of unicasts.  However, there's already a loop in the
generic send_IPI_mask code to do that ...  there's no need to call
send_IPI_mask once for each CPU.  The comment I wrote at the time even
noted that this was silly.