Commit 03ffc4d6 added logic to bypass all caching behavior in
LookupOpclassInfo when CLOBBER_CACHE_ALWAYS is enabled.  It doesn't
look like I stopped to think much about what that would cost, but
recent investigation shows that the cost is enormous: it roughly
doubles the time needed for cache-clobber test runs.

There does seem to be value in this behavior when trying to test
the opclass-cache loading logic itself, but for other purposes the
cost is excessive.  Hence, let's back off to doing this only when
debug_invalidate_system_caches_always is at least 3; or in older
branches, when CLOBBER_CACHE_RECURSIVELY is defined.

While here, clean up some other minor issues in LookupOpclassInfo.
Re-order the code so we aren't left with broken cache entries (leading
to later core dumps) in the unlikely case that we suffer OOM while
trying to allocate space for a new entry.  (That seems to be my
oversight in 03ffc4d6.)  Also, in >= v13, stop allocating one array
entry too many.  That's evidently left over from sloppy reversion in
851b14b0.

Back-patch to all supported branches, mainly to reduce the runtime
of cache-clobbering buildfarm animals.

Discussion: https://postgr.es/m/1370856.1625428625@sss.pgh.pa.us