Move FLOAT_CMP upstream and/or combine with numpy/scipy?

[mhvk]

As noted originally in astropy/astropy#6554, numpy was thinking about their own version of something like FLOAT_CMP and ended up adapting scipy's refguide [1] for its docstring testing. It includes float comparisons and "numbers may vary". We should check whether we can also use this, to make the ecosystem a bit more uniform. In this respect, probably good to remember that this was put also to python-ideas by @kdavies4 with Coghlan concluding it was a good idea [2]. (See also numpy/numpy#9415 (comment))

[1] https://github.com/scipy/scipy/blob/master/tools/refguide_check.py
[2] https://mail.python.org/pipermail/python-ideas/2014-August/028559.html

[drdavella]

I'd be in favor, especially given the fact that we recently discovered some issues with the way that FLOAT_CMP interacts with other doctest flags. See #54.

[hameerabbasi]

This is something that would be very useful. I can look into it, definitely. 😄

[nstarman]

Pytest has added NUMBER (see https://docs.pytest.org/en/6.2.x/doctest.html?highlight=doctest#using-doctest-options). Maybe FLOAT_CMP can just be a thin redirect to that?

[pllim]

I think we decided that NUMBER isn't as flexible as we need it to be (we support atol and rtol here, but it does not).

Superseded by #199

[mhvk]

Let's keep this open for the moment, since perhaps NUMBER is really good enough: its atol is just set by the last digit, so adjustable in that sense.