Original commit was cmusphinx/sphinxbase@e59cac4

The resulting FSGs are, yes, quite bad, but the optimization was incorrect, and the
proposed fix for it did not actually fix the problem, but made the code more
complex.  In general I think the Knuth quote about premature optimization is
just an excuse for people to write bad code, but in this case it seems pertinent

The change I reverted was *partially* correct as it noted that a right-recursive rule
cannot reach the final state on its own, so put this change back

with the side effect that it verifies the correctness of the
tests in test-sphinx_jsgf2fsg.sh