lazy vxid locks, v3

I took another look at v2 of my lazy vxid locks patch and realized
that it was pretty flaky in a couple of different ways.  Here's a
version that I think is a bit more robust, but considering the extent
of the revisions, it probably needs another round of review from
someone before I commit it.

Any review appreciated; I would prefer not to have to wait until
October to get this committed, since there is quite a bit of follow-on
work that I would like to do as well.  FWIW, the performance
characteristics are basically identical to the previous versions,
AFAICT.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company