Refactoring SysCacheGetAttr to know when attr cannot be NULL

Today we have two fairly common patterns around extracting an attr from a
cached tuple:

  a = SysCacheGetAttr(OID, tuple, Anum_pg_foo_bar, &isnull);
  Assert(!isnull);

  a = SysCacheGetAttr(OID, tuple, Anum_pg_foo_bar, &isnull);
  if (isnull)
    elog(ERROR, "..");

The error message in the elog() cases also vary quite a lot.  I've been unable
to find much in terms of guidelines for when to use en elog or an Assert, with
the likelyhood of a NULL value seemingly being the guiding principle (but not
in all cases IIUC).

The attached refactoring introduce SysCacheGetAttrNotNull as a wrapper around
SysCacheGetAttr where a NULL value triggers an elog().  This removes a lot of
boilerplate error handling which IMO leads to increased readability as the
error handling *in these cases* don't add much (there are other cases where
checking isnull does a lot of valuable work of course).  Personally I much
prefer the error-out automatically style of APIs like how palloc saves a ton of
checking the returned allocation for null, this aims at providing a similar
abstraction.

This will reduce granularity of error messages, and as the patch sits now it
does so a lot since the message is left to work on - I wanted to see if this
was at all seen as a net positive before spending time on that part.  I chose
an elog since I as a user would prefer to hit an elog instead of a silent keep
going with an assert, this is of course debateable.

Thoughts?

--
Daniel Gustafsson