genrecog: warn of unmatchable predicate codes

[DJ Delorie]

> ALL of them are incorrect conversions TO define_predicate, and thus
> regressions from 3.4.  They MUST be fixed.

Hmmm... interesting way of looking at it.  I was thinking in terms of
testsuite results, and who knows how changing a predicate would affect
testsuite results.  I'm certainly willing to make this an error, but
it's impractical for me to be responsible for testing all the
platforms for testsuite regressions caused by any predicate fixings.
I could probably write a perl script to scan all the .md files for
such errors and blindly fix them, but who knows what effects that will
cause.