A concrete proposal for NaNs in = and <.  As mentioned previously,
mpz_cmp_d doesn't accept them.

        * numbers.c (scm_num_eq_p, scm_less_p): Don't pass NaN to mpz_cmp_d.

        * tests/numbers.test (=, <): Add tests involving NaNs.

Are the comments correct about SCM_COMPLEXP numbers never having a
zero imaginary part?  I suspect this means the mpz_cmp_d's in the
complex/bignum cases are not reached.  In fact perhaps any comparison
(= complex non-complex) would be #f without any testing.