Marten van Kerkwijk

OK, always going through SI will at least be consistent, though I think it will also lead to confusion. But probably unsolvable... Though with the IAU definitions of those types...

@nstarman - I don't quite remember the details, but seem to recall I've encountered something similar before. I think one issue is that even with `dtype=object`, numpy will try to...

That last error is raised in `Table.__array__` -- I guess we could change that to look at `dtype` and if it is `object` return a scalar array with the table...

@pllim - sorry to bug, but do you happen to know what is up with `readthedocs`? Do we need to rebase or so?

Note that we could avoid repeated validation also by memoizing the outcome. Though I find the logic that is being introduced here, or doing validation on setting, and broadcasting later,...

Hmm, yes, getting a secondary frame attribute may in fact be a reason for validating on `__get__` in the first place.

For the secondary attribute, I'm not sure we have a test... Right now, it seems we only use it in `FK4`, for ``` obstime = TimeAttribute(default=None, secondary_attribute='equinox') ``` validation for...

I don't think we should just deprecate the download & cache behaviour - it would often be what users need for repeatedly running a script to analyze something, etc. Note...

p.s. An argument against always using `fsspec` is that then the behaviour of astropy depends on what packages are installed.

@maxnoe - in principle OK, but I think the logic should be reversed: the frames do not really know about `SkyCoord`, while `SkyCoord` knows about frames, so my sense would...