Closes gh-19171

gh-19171 reported that numpy.testing assertion messages sometimes referred to the actual and desired results as x and y instead of ACTUAL and DESIRED.

gh-20161 appeared to fix this issue and seemed close to being merged, but it became inactive.

This PR squashes the commits from the original PR and resolves merge conflicts. I also made additional changes required to improve consistency.

Here is the state of things:

• assert_array_almost_equal_nulp and assert_array_max_ulp are symmetric with respect to the order of the inputs. The failure messages do not refer to the arguments by name; e.g. "Arrays are not equal to 1 ULP". (Previously, one of these referred to the arguments by name; the other did not.)
• assert_array_less does not have a notion of "actual" and "desired", but the order of the inputs matters. The arguments are named x and y, so they are referred to as such in the failure message; e.g. "Arrays are not strictly ordered x < y". (Previously, the message was less specific: "Arrays are not less-ordered".)
• assert_allclose, assert_equal, assert_almost_equal, assert_approx_equal have arguments named actual and desired; in the failure message, the arguments are now always referred to as ACTUAL and DESIRED. (Previously, they were sometimes referred to as x and y, which was the original issue.)
• assert_array_almost_equal and assert_array_equal have arguments named x and y. In the error message, they are now always referred to as ACTUAL and DESIRED. (Previously, they were always referrred to as x and y, but this is not as bad as the original issue because the docstrings refer to these arguments as the "actual" and "desired" objects. )

Regarding the last point - the original issue was that the arguments of assert_equal were not referred to by their names; in some sense, this problem has just been transferred to assert_array_equal and assert_array_almost_equal.

To fully resolve this, I'd like to change the names of the first two arguments of assert_array_equal and assert_array_almost_equal from x and y to actual and desired. Of course, they would still support keyword arguments x and y to preserve backward compatibility, but changing the names that show in the docstring would make the asymmetry of assert_array_equal more obvious and reduce the potential confusion of referring to x as ACTUAL and y as DESIRED in the failure message.

The decorator we use in SciPy for this purpose, _rename_parameter, would make this easy. Shall I do this in a follow-up PR? (Does NumPy have its own decorator for this purpose, and if not, should I vendor SciPy's decorator?)