Proper `epsilon` for complex?

It seems that the finite difference routines would "just work" if `epsilon` was changed to the "correct value".
Unless there's some kind of norm argument, I would propose changing each part using the current epsilon. For example, if `epsilon1` is from `real(x)` and `epsilon2` is from `imag(x)`, then `epsilon = epsilon1 + im*epsilon2`.

Would this be fine? This definition would make it safe for the case where a user is just "packing" floats into complex numbers, which can be common in physical applications. If this is acceptable I'll make a PR.