Currently, the mask of a MaskedArray is incorrectly changed in the following:

import numpy as np; from numpy import isclose, nan
m = np.ma.MaskedArray([nan, nan], mask=[True, False])
isclose(m, m, equal_nan=True)

yields

masked_array(data = [True True],
             mask = [False False],
       fill_value = True)

The problem lies in the fact that in isclose, the following final assignment is done

cond[isnan(x) & isnan(y)] = True 

The problem here is that the mask of the index is ignored in __setitem__ (and it would be very hard to do otherwise), but and hence all elements are being set. Since True has no mask, the mask of all ements is cleared. To avoid this, this PR ensured that the vallues used include a mask, by replacing the above line with

both_nan = isnan(x) & isnan(y)
cond[both_nan] = both_nan[both_nan]

A test case is included as well.

p.s. This bug was found while trying to allow MaskedArray to behave better with subclasses; see #3907