Negative zero not preserved

[tomwhite]

This was picked up in the array API tests: https://github.com/cubed-dev/cubed/actions/runs/15697743690/job/44226032553

AssertionError: out=0.0, but should be -0 [sin()]

Reproducer:

>>> import numpy as np
>>> a = np.array([0.0, -0.0])
>>> np.sin(a)
array([ 0., -0.])
>>> import cubed as xp
>>> b = xp.asarray(a)
>>> xp.sin(b).compute()
array([0., 0.])

With Zarr 2 it works:

>>> import numpy as np
>>> a = np.array([0.0, -0.0])
>>> np.sin(a)
array([ 0., -0.])
>>> import cubed as xp
>>> b = xp.asarray(a)
>>> xp.sin(b).compute()
array([ 0., -0.])

[tomwhite]

Opened zarr-developers/zarr-python#3144

[tomwhite]

I worked around this by setting write_empty_chunks to True (#733), which we want to do in Cubed anyway to avoid comparing every chunk to the fill value. The upstream issue is still valid though.