Finally tracked down problems in regressions and fuzz results (this also

Finally tracked down problems in regressions and fuzz results (this also
explains why my build was giving different answers from Dejan's build).
Problem was that if you run:

cvc4 --check-models foo.smt

it would fail, but if you run

cvc4 --check-models --produce-models foo.smt

it would succeed.  Even though produce-models is supposed to be automatically
set when you set check-models, it seems this was happening too late.  Fixed by
removing any distinction between produce-models and check-models in the
heuristic setting code.  Kind of ugly though.

Also, disable models if nonlinear arithmetic is on.