[PATCH] pg_upgrade: Split off pg_fatal() from pg_log()

The experiences with elog() and ereport() have shown that having one
function that can return or not depending on some log level parameter
isn't a good idea when you want to communicate well with the compiler.
In pg_upgrade, there is a similar case with the pg_log() function.
Since that isn't a public API, I'm proposing to change it and introduce
a separate function pg_fatal() for those cases where the calls don't
return.