Doing NMI startup as an early initcall doesn't work because we need
to have SMP started up by then.

So we'd only NMI startup one cpu, which causes perf PMU grab to
BUG because the nmi_active count isn't what it's supposed to be.

This also points out that we don't have proper CPU up/down notifiers
for the NMI code which will need to be fixed at some point.
Signed-off-by: David S. Miller <davem@davemloft.net>