Improve accuracy of expm1 for complex arguments

[thangleiter]

Fixes #418.

[thangleiter]

The imaginary part of the result differs slightly from the numpy result (hence the nulp=2500) in the test. Strangely, this difference disappears if the exponential of the real part is assigned to a temporary value (i.e., as it is currently computed):

static void
nc_expm1(npy_cdouble *x, npy_cdouble *r)
{
    double a = sin(x->imag / 2);
    double b = exp(x->real);
    r->real = expm1(x->real) * cos(x->imag) - 2 * a * a;
    r->imag = b * sin(x->imag);
    return;
}

Any idea why @robbmcleod?

[robbmcleod]

Closing as I implemented myself, thanks for the contribution though!