Some ERFA ufuncs do not work with float32

I found one particular function in ERFA which does not accept float32 values, although there are probably more:

In [3]: xyz = np.array([[1, 0, 0], [0.9, 0.1, 0]])

In [4]: _erfa.p2s(xyz)
Out[4]: 
(array([0.        , 0.11065722]),
 array([0., 0.]),
 array([1.        , 0.90553851]))

In [5]: _erfa.p2s(xyz.astype(np.float32))
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-5-cbd078a35d7c> in <module>()
----> 1 _erfa.p2s(xyz.astype(np.float32))

~/.miniconda36/envs/py37/lib/python3.7/site-packages/astropy/_erfa/core.py in p2s(p)
  18187 
  18188     """
> 18189     theta, phi, r = ufunc.p2s(p)
  18190     return theta, phi, r
  18191 

RuntimeError: no relevant input found; should not happen!

This issue manifests itself in quite strange ways:

In [97]: icrs = ICRS(CartesianRepresentation(xyz.astype(np.float32), xyz_axis=-1, unit="km"))

In [98]: q = icrs.spherical
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-98-73fe0562a4ab> in <module>()
----> 1 q = icrs.spherical

~/.miniconda36/envs/py37/lib/python3.7/site-packages/astropy/coordinates/baseframe.py in spherical(self)
   1661         # TODO: if representations are updated to use a full transform graph,
   1662         #       the representation aliases should not be hard-coded like this
-> 1663         return self.represent_as('spherical', in_frame_units=True)
   1664 
   1665     @property

~/.miniconda36/envs/py37/lib/python3.7/site-packages/astropy/coordinates/baseframe.py in represent_as(self, base, s, in_frame_units)
   1072                 diff = data.differentials['s']  # TODO: assumes velocity
   1073             else:
-> 1074                 data = self.data.represent_as(representation_cls)
   1075 
   1076             # If the new representation is known to this frame and has a defined

~/.miniconda36/envs/py37/lib/python3.7/site-packages/astropy/coordinates/representation.py in represent_as(self, other_class, differential_class)
    688             if other_class is not self.__class__:
    689                 # The default is to convert via cartesian coordinates
--> 690                 new_rep = other_class.from_cartesian(self.to_cartesian())
    691             else:
    692                 new_rep = self

~/.miniconda36/envs/py37/lib/python3.7/site-packages/astropy/coordinates/representation.py in from_cartesian(cls, cart)
   1664         p = cart.get_xyz(xyz_axis=-1)
   1665         # erfa p2s: P-vector to spherical polar coordinates.
-> 1666         return cls(*erfa_ufunc.p2s(p), copy=False)
   1667 
   1668     def norm(self):

~/.miniconda36/envs/py37/lib/python3.7/site-packages/astropy/units/quantity.py in __array_ufunc__(self, function, method, *inputs, **kwargs)
    461 
    462         # Call our superclass's __array_ufunc__
--> 463         result = super().__array_ufunc__(function, method, *arrays, **kwargs)
    464         # If unit is None, a plain array is expected (e.g., comparisons), which
    465         # means we're done.

RuntimeError: no relevant input found; should not happen!

which comes from here:

https://github.com/astropy/astropy/blob/9bffef18249f802792a81800a44efc038ca3ac4c/astropy/_erfa/ufunc.c.templ#L625-L626

And, in my case, was difficult to spot because iterating by value did work, see #8614.

[astrojuanlu]

Checked, thanks @mhvk @pllim! 👍

[pllim]

@Juanlu001 and @kyleaoman , now that #8759 is in master, please check if your use case now works with the latest master. Thanks!