Represent None component.units as empty strings

[dhomeier]

Description

To fix issues as in spacetelescope/jdaviz#1909 where f'{unit}' returns unexpected results, this returns self._units of None as '' for all Component subclasses. This is analogous to how these are already handled in Coordinates.
In case the internal value of self._units == None is still required somewhere, the second commit reverses the transformation on setting units, but it does not seem to be required for passing tests.

[pllim]

What was the design decision to use None in the first place? astropy.units would happily parse '' into u.dimensionless_unscaled.

[astrofrog]

I think I'm happy for the public API to not change, I just wanted to be be able to distinguish internally between units being set vs default.

[pllim]

Is this related?

DeprecationWarning: Passing np.nan to mean no clipping in np.clip has always been unreliable, and is now deprecated. In future, this will always return nan, like it already does when min or max are arrays that contain nan. To skip a bound, pass either None or an np.inf of an appropriate sign.

[dhomeier]

I think I'm happy for the public API to not change, I just wanted to be be able to distinguish internally between units being set vs default.

Then perhaps it's better to cherry-pick the first commit, so passing units='' would not be translated to None.

[dhomeier]

Is this related?

DeprecationWarning: Passing np.nan to mean no clipping in np.clip has always been unreliable, and is now deprecated. I

To the units treatment? No, showed already up before the units PR was merged; I suspect it is related to the

glue/glue/viewers/scatter/layer_artist.py:60: RuntimeWarning: All-NaN slice encountered
return 10. ** (np.log10(np.nanmax(array)) * self.contrast)

from the same test_density_map_incompatible_subset, but can't tell where exactly this is raised.

[astrofrog]

Then perhaps it's better to cherry-pick the first commit, so passing units='' would not be translated to None.

@dhomeier - sounds good, I'll remove the second commit

[astrofrog]

The other jdaviz failures are unrelated to this PR, so merging this and will open a separate PR to fix the other issue.