Numpy dev array printing - update all tests to new numpy style - #5423

@taldcroft, @astrofrog, @adrn: this PR makes tests pass on numpy-dev, but at the price of changing how coordinate output looks, to match what the new numpy array2print does for structured arrays. I think it it the right way forward, but would like to be sure there is agreement.

Just to be a bit clearer about the situation: With current master, you'll get different-looking output depending on whether you use a numpy release or the development version; with this PR, the output is always like that with the development version. To get a sense of how things change:

import astropy.units as u
from astropy.coordinates import SphericalRepresentation
r1 = SphericalRepresentation(lon=1 * u.deg, lat=2.5 * u.deg, distance=1 * u.kpc)
r2 = SphericalRepresentation(lon=[1,20,-3, -3, 2, 1] * u.deg, lat=[2.5,0,-2.5, -2.5, 1, 1] * u.deg, distance=1 * u.kpc)
r1
# numpy <1.12, current master
<SphericalRepresentation (lon, lat, distance) in (deg, deg, kpc)
    (1.0, 2.5, 1.0)>
# numpy dev current master/this PR any numpy
<SphericalRepresentation (lon, lat, distance) in (deg, deg, kpc)
    ( 1.,  2.5,  1.)>
#
r2
# numpy <1.12, current master
<SphericalRepresentation (lon, lat, distance) in (deg, deg, kpc)
    [(1.0, 2.5, 1.0), (20.0, 0.0, 1.0), (357.0, -2.5, 1.0),
     (357.0, -2.5, 1.0), (2.0, 1.0, 1.0), (1.0, 1.0, 1.0)]>
# numpy-dev, current master/this PR, any numpy
<SphericalRepresentation (lon, lat, distance) in (deg, deg, kpc)
    [(   1.,  2.5,  1.), (  20.,  0. ,  1.), ( 357., -2.5,  1.),
     ( 357., -2.5,  1.), (   2.,  1. ,  1.), (   1.,  1. ,  1.)]>

You see that three things happened: trailing 0 are omitted; there is always either a space or - sign; numbers are aligned nicely vertically.

What do you think? Is the change for the better? If you look at representation.py, you'll see that there was some code there in __repr__ to try to mimic numpy array printing, but it was incomplete. So, in that sense, I think this PR just does what was originally intended.

Now there is an alternative that is worth considering: should we really care how output looks at all, as long as numbers are actually present (and correct, of course)? If so, we should remove or rewrite the tests that fail with current master and numpy-dev, and also find a way for doctests to work irrespective of formatting. Overall, I felt changing the tests made more sense, since I do not think this is going to change again soon...

@mhvk - I prefer the new-style formatting overall and definitely vote for consistency across numpy versions. So 👍 from me.

@mhvk I think the new formatting looks much cleaner. I also agree that we should just check for correctness and not formatting in the tests!

(Rebased following merger of #5254)

The one remaining failure with numpy-dev is unrelated to this PR (and on its way to being fixed; numpy/numpy#8218). Since there was general agreement on this change in printing, I'll merge this now.

This is exceedingly difficult to backport do to conflicts almost everywhere, so re-classifying to 1.3.x . We can re-visit this if we feel we need 1.0.x tests with numpy-dev, but my sense is that this is not critical...?

Re: Backport and numpy-dev. IMHO, whoever still using 1.0.x is unlikely to be using numpy-dev, so my vote is "not critical."

I agree with @pllim here. Whomever is flexible enough to upgrade to numpy 1.12, should be flexible enough to upgrade to astropy 1.3.

Yes, agreed, especially as it only affects running the tests, not using actual code.