Shortened display of NumPy arrays in DataArray.__repr__

[shoyer]

Fixes #680

This shortens the array portion of the DataArray repr to reasonable lengths for high dimensional arrays:

Array	Old length	New length
np.random.randn(500)	167	1
np.random.randn(20, 20)	150	7
np.random.randn(5, 10, 15)	254	24
np.random.randn(5, 10, 15, 3)	248	19
air_temperature['air']	41	24
rasm['Tair']	37	24

The last two entries refer to the corresponding datasets loaded from xarray.tutorial.load_dataset.

Importantly, this change does not preclude future improvements, such as those suggested in #1044.

Example of why I switched to lower precision:

In [24]: x = np.random.randn(20, 100)

In [25]: x
Out[25]:
array([[-0.11684307,  0.67454961, -0.0617408 , ..., -0.08199539,
         0.11447149,  1.18730556],
       [ 1.17061389, -0.72944301, -0.45572465, ...,  1.4337907 ,
        -0.80810471,  0.51210789],
       [-0.69365824, -1.1452782 ,  1.8398729 , ...,  0.06559557,
         1.77739577,  2.12243584],
       ...,
       [-2.52712434,  0.53711727, -0.08554198, ...,  0.79291491,
        -0.4017885 , -0.58840279],
       [ 0.12926157, -0.71417823, -0.40803449, ...,  0.97339737,
         1.58154386, -0.78457038],
       [ 1.71986036, -1.02348818,  1.79216626, ..., -0.11309842,
        -1.00089845,  0.11611589]])

In [27]: print(xr.core.formatting.short_array_repr(x))
array([[-0.116843,  0.67455 , -0.061741, ..., -0.081995,  0.114471,  1.187306],
       [ 1.170614, -0.729443, -0.455725, ...,  1.433791, -0.808105,  0.512108],
       [-0.693658, -1.145278,  1.839873, ...,  0.065596,  1.777396,  2.122436],
       ...,
       [-2.527124,  0.537117, -0.085542, ...,  0.792915, -0.401788, -0.588403],
       [ 0.129262, -0.714178, -0.408034, ...,  0.973397,  1.581544, -0.78457 ],
       [ 1.71986 , -1.023488,  1.792166, ..., -0.113098, -1.000898,  0.116116]])

[shoyer]

Here is what @MaximilianR's example from #680 looks like with this change:

In [31]: print(xr.core.formatting.short_array_repr(np.nan * np.zeros((2, 2, 5, 6771, 58))))
array([[[[[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]],

         ...,
         [[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]]],


        [[[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]],

         ...,
         [[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]]]],



       [[[[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]],

         ...,
         [[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]]],


        [[[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]],

         ...,
         [[ nan, ...,  nan],
          ...,
          [ nan, ...,  nan]]]]])

It's not ideal (still pretty hard to make sense of the number of dimensions, and lots of wasted horizontal space), but it's definitely an improvement.

[shoyer]

OK, merging this for v0.9