Regression with Astropy 7.2: WCSLink does not raise on WCS without physical_types

[dhomeier]

Describe the bug
Failure of test_wcs_autolink_emptywcs in all dev jobs, can be tracked down to wcs_autolink returning a non-empty list with one link.

To Reproduce
Test glue/plugins/wcs_autolinking/tests/test_wcs_autolinking.py or manually run WCSLink(data1, data2) with the data as setup there.

Expected behavior
WCSLink should raise in

glue/glue/plugins/wcs_autolinking/wcs_autolinking.py

Lines 210 to 211 in bf65166

	if pixel_cids1 is None:
	raise IncompatibleWCS("Can't create WCS link between {0} and {1}".format(data1.label, data2.label))

but returns a link with link._physical_types_* = [None] instead.
Might need to check explicitly for valid _physical_types_ there in addition to pixel_cids1, but perhaps should investigate the responsible change in astropy.wcs first.

Details:

• Operating System: macOS 11.7.10 or any in -dev job, e.g. in
  #2547 CI

[CyclingNinja]

I've tracked the initial bad commit to which was part of broken world to pixel slicing fix a week or two ago on this PR astropy/astropy#18394

baaf33ff2fd2d6a8e91f9971400a210b3844c3d6

[dhomeier]

Thanks for identifying the change @CyclingNinja! It looks like WCSLink previously has been failing on

glue/glue/plugins/wcs_autolinking/wcs_autolinking.py

Lines 79 to 82 in bf65166

	def get_cids_and_functions(wcs1, wcs2, pixel_cids1, pixel_cids2):
	def forwards(*pixel_input):
	return pixel_to_pixel(wcs1, wcs2, *pixel_input)

raising a TypeError: iteration over a 0-d array on the empty (NAXIS=0) wcs, but that part now passes after the astropy fix changed the pixel_to_world_values return to a length-1 list/array.
So the whole check for WCS compatibility is relying on their dimensionality rather than actual physical types; should that be reconsidered?

[CyclingNinja]

PR to glue-core to address this https://github.com/glue-viz/glue/pull/2551/files