A recent change to calls to map[io]dev that ommitted the use
of a bustype seem to leave one with a few nodes that have a
value of zero for the rr_iospace tag portion of register
addresses.

This isn't quite enough of a hint to pmap_enter to map the pages
as part of I/O space- bad things happened.

I've hacked up sbus.c so I can proceed with stuff for myself, but
I'm not in tune enough with the current port-sparc effort to really
know where the right place to fix this should be- whether it should
be the place that fills in romaux info from the prom, or whether it
should always be the parent node (e.g., sbus always sets iospace
to children nodes to 1, a vme parent node knows what to do for
it's children, etc...). This also begs the question about whether
NVRAM boards which present their memory as registers should be mapped
non-cached, etc....