fix: xarray roi selection

[tmichela]

This fixes a few bugs found after #662 was deployed:

For example, with Jungfrau data, shape (1,512,1024) per trainId

• Ellipsis
  • np.s_[:1,30:40,...]: Returns dim_1 with coords arange(30,40) only. dim_2 has no coords
  • np.s_[...,30:40]: CoordinateValidationError
• Single valued dimensions are omitted from coords/dimension names
  • np.s_[0,30:40,:]: Returns array with dim_0 arange(30,40), which should be dim_1, retaining dim_0=0 as collapsed dimension
  • np.s_[0, [30, 33], :]: CoordinateValidationError: conflicting sizes for dimension 'dim_0'
• Related to single valued dimensions, this no longer works:
  • roi = np.s_[0]: returns TypeError: 'int' object is not iterable
  • np.s_[0, 10:20,30:40] with extra_dims=['a', 'b', 'c']: ValueError: different number of dimensions on data and dims: 3 vs 4

Enhancement:
Provides a new extra_coords argument that takes either a dict for user provided coordinates, or a bool to generate default coordinates.

[tmichela]

@takluyver as discussed offline, we're now adding coordinates if we provide extra_dims argument or if the new extra_coords is True. extra_coords can also be used to provide custom coordinates as a dict.

[takluyver]

This can also be simplified now that we're not supporting None

[tmichela]

updated in edc7554

[takluyver]

So if you pass both a coordinates dict and a ROI, the coordinates should describe the data after selecting the ROI, if I've understood correctly? I don't have a feeling about which way is better, but it's probably good to document it.

[tmichela]

updated in edc7554

[takluyver]

So just to check I've understood correctly, passing only roi= will no longer attach any coordinates (besides train IDs), and you'll need to use extra_dims or extra_coords to have automatic integer coordinates?

I think this design sounds OK.

[takluyver]

In terms of the implementation, I think you can give xarray extra coordinates for dimensions that aren't in the data, without having to do the trick with expanding the array dimensions and then squeezing them back out. Have I misunderstood something? Does that produce a subtly different result?

E.g. attaching a 'z' coordinate with no z dimension:

xr.DataArray(
    np.random.randn(2, 3),
    dims=("x", "y"),
    coords={"x": [10, 20], 'z': 6}
)

[tmichela]

So just to check I've understood correctly, passing only roi= will no longer attach any coordinates (besides train IDs), and you'll need to use extra_dims or extra_coords to have automatic integer coordinates?

Yes, that's exactly that!

[takluyver]

Thanks, LGTM!

[tmichela]

Thanks for the review!