After considerable thinking, I have come to the conclusion that this
assert is bogus.

There is a window from our lookup until we lock the refcount where
a different thread conceiveably could loose the last reference to
the objhdr, before we get the lock to increase it.

The situation the assert was meant to protect, where we insert 
a new objhdr, is handled in the "oh == noh" clause earlier, were
the weaker "> 0" check is used (rather than "== 1") to cater for
the exact same race.

Fixes: #761
Fixes: #783



git-svn-id: http://www.varnish-cache.org/svn/branches/2.1@5368 d4fa192b-c00b-0410-8231-f00ffab90ce4