When the LUA floating point model is INT64, we don't need to do the
overflow dance. This avoids compile errors on latter-day gcc compilers
as well as simplifies the generated code.

Reviewed by: kevans@