Have coordinate transformations preserve (float?) dtypes

This may be more a part of units/Quantity than coordinates, but the proximate problem is coordinates so I’m starting there… the core problem is this:

>>> import numpy as np
>>> from astropy import units as u
>>> from astropy import coordinates as coo
>>> icrs = coo.ICRS(coo.CartesianRepresentation(np.array([1,2,3], dtype='float32')*u.kpc))
>>> icrs.data.xyz.dtype
dtype('float32')
>>> icrs.transform_to(coo.Galactic).data.xyz.dtype
dtype('float64')

that is, if I start with a dtype that’s float32 the transformation process turns it into float64. It’s not clear we formally expected it to stay the same, but it would be useful for some applications of the dtype is preserved…

cc @adrn

[mhvk]

I agree that for some erfa routines it really should not be difficult - and that particularly includes the transformation routines we need for the representations. I think we’d want to use particular numpy types rather than C types (npy_double), etc., but that is really an implementation detail - it will just set us up more easily for when numpy does get a proper float128…

p.s. Of course, for many other erfa routines, like the time conversion ones, it makes no sense, as those really are written for doubles.

[mhvk]

Either, really, but it is a generally true for ufuncs that they allow “same_kind” casting. In principle, we could avoid this by adding casting='safe', though given that we do not have full precision routines, it seems rather pointless.