existence of np.float128 is platform-dependent. since it always just an alias to np.longdouble we should use that instead.

numpy.float128 is not guaranteed to exist on all systems, but numpy.longdouble is. Since numpy.float128 is just an alias, we can move to using np.longdouble without a loss of precision.

The same is true for numpy.complex256 and numpy.clongdouble. Both are addressed in my fix/np-dtypes branch.