This patch moves the per-CPU MCA data allocation out of per_cpu_init()
so the code can be shared for contig and discontig memory
architectures.  Also, it means we can revert back to the old way
of doing per_cpu_init() on UP.

Also cleanup initialization of ar.k3 in setup.c.  Need to use
ia64_tpa() rather than __pa() because address will be in region 5
on UP build.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>