The corresponding relminxid patch; try 1

Hi,

This is the relminxid patch corresponding to the pg_ntclass patch I just
posted.  Obviously, the relminxid and relvacuumxid fields are in
pg_ntclass (not pg_class like in the previous incarnations of this
patch).  This makes the whole business much saner and now I don't need
to insert bogus CommandCounterIncrement calls.  Regression tests pass.

The thing that bothers me most about this is that it turns LockRelation
into an operation that needs to heap_fetch() from pg_ntclass in some
cases, and possibly update it.  I think we should consider some sort of
"non-transactional shared cache" for storing RELKIND_NON_TRANSACTIONAL
catalog entries.  Eventually it may help the sequences stuff as well, if
we implement sequences using that kind of catalog.

The documentation changes may be a bit off in this patch, since I didn't
worry about merging it with the pg_ntclass patch.  But it's easy to
correct and I'll do it before committing it.

My intention is to wait two or three days after committing the
pg_ntclass patch, and then commit this one, unless I hear objections
before that.