BUG: arctan inaccurate for small complex numbers

Consider the following examples:
The first few terms of taylor-series for arctan is given by

>>> import numpy as np
>>> def arctan_3(x):
...     return x-x**3/3+x**5/5

Then we compare the taylor-series with numpy.arctan for small complex number:

>>> x =  0.01+1je-14
>>> v1 = np.arctan(x); v1
(0.0099996666866652394+9.9920072216263095e-15j)

>>> v2 = arctan_3(x); v2
(0.009999666686666667+9.9990001e-15j)


>>> (v2-v1).imag/v2.imag
0.0006993577661521059

Lets compute the complex step derivative of arctan

>>> y = 0.01
>>> h = 1e-14
>>> np.arctan(y+1j*h).imag/h
0.99920072216263089

The exact value is

>>>1./(1+y**2) 
0.9999000099990001:

>>> arctan_3(1j*h).imag/h
0.9999000100000001

In both the examples the relative error of the complex part of numpy.arcsin function is around 1e-4.

It should be relatively easy to fix this as the complex part of arctan function is proportional to x when x is small.

Reproduce the reported cases with Python and numpy.arctan, comparing the complex component against the Taylor-series results and the exact derivative. Trace the arctan implementation and its numerical tests; done means small complex inputs produce accurate imaginary parts without regressing existing arctan behavior.