- Always enable __array_function__ overrides.
- Remove special cases for Python 2 compatibility.
- Document these changes in 1.17.0-notes.rst.

It will be good to see ASV numbers to understand the performance implications
of these changes. If need be, we can speed up NumPy functions internally by
using non-dispatched functions (with ``.__wrapped__``).

Example behavior:

>>> x = np.array([1, 2, 3])
>>> y = np.array([1, 2, 3.0001])
>>> np.testing.assert_allclose(x, y)
AssertionError:
Not equal to tolerance rtol=1e-07, atol=0

Mismatch: 33.3%
Max absolute difference: 0.0001
Max relative difference: 3.33322223e-05
 x: array([1, 2, 3])
 y: array([1.    , 2.    , 3.0001])

Motivation: when writing numerical algorithms, I frequently find myself
experimenting to pick the right value of `atol` and `rtol` for
`np.testing.assert_allclose()`. If I make the tolerance too generous, I risk
missing regressions in accuracy, so I usually try to pick the smallest values
for which tests pass. This change immediately reveals appropriate values to
use for these parameters, so I don't need to guess and check.

Most of this was already done, but we were still raising an error for
declaration after a statement because the Windows Python 2.7 compiler
did not allow it. We can fix this now as NumPy >= 1.17 has dropped
Python 2.7 support.

MAINT: Update changelog.py for Python 3.

- Don't use UTF8Writer in Python 3.
- Don't use print() for blank lines.

This is left compatible with Python 2 as someone may use it with
that version and there is no need to force the change.

* added lib.recfunctions to refguide
PUBLIC_SUBMODULES, as the doctests
were otherwise not getting executed

* fixed a failing doctest in
recfunctions after above activation