The intention was to catch functional regressions in non-matching
functions but this doesn't actually work well because referenced
code and data may have moved after a function was dumped, in which
case the reference checker is unable to verify that function calls
and data references are still correct.

Considering this increases the amount of complexity in the checker
and makes the workflow more complicated for contributors, let's just
drop this mechanism. It isn't worth it.

Putting binary files in the repo is also pretty meh.

It won't be accurate because the factory functions can and will be
automatically generated.

This makes it possible to catch regressions for non-matching functions,
especially those that only have minor issues.

This also reclassifies some minor non-matchings as major non-matchings
whenever it's really not obvious to see that they are equivalent.

And make the output clearer.