intel-iommu: use atomic ops for irte index alloc

To allow concurrent allocation of irte index. Meanwhile, move the IRTE
setup debug line into the alloc since vtd_setup_msi() might not be the
only one to dump this info in the future.

Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>