Item handler: fall back to Zarr asset or raise a clear error (#68)

[ghostiee-11]

What

When xpystac.to_xarray(item) is called on a pystac.Item that has no kerchunk reference, the existing handler runs off the end of the function and returns None. That None then propagates up to xarray's backend, which dereferences dataset.variables and raises an opaque AttributeError: 'NoneType' object has no attribute 'variables' from deep inside _protect_dataset_variables_inplace.

Reproduced on the live Earth Search NAIP catalog and in the Lumen STACSource path (holoviz/lumen#1867):

import pystac_client, xarray as xr, xpystac

client = pystac_client.Client.open('https://earth-search.aws.element84.com/v1')
item = next(client.search(collections=['naip'], max_items=1).items())
xr.open_dataset(item, engine='stac')
# AttributeError: 'NoneType' object has no attribute 'variables'

Fix

Add an explicit fallback in the Item handler:

• If the Item has any Zarr / icechunk asset, dispatch to the registered Asset handler with that asset (so an item that ships a dataset-level asset alongside other formats now opens cleanly).
• Otherwise, raise a ValueError naming the observed media_types and, if the item has a COG asset (the common case), pointing the user at stackstac / odc-stac, which are the right tools for COG-only items.

After the fix, the same snippet returns a clean error:

ValueError: Item 'me_m_4506963_se_19_030_20231115' has no Zarr or kerchunk
asset that xpystac can open as a dataset; observed asset media_types:
['application/xml', 'image/tiff; application=geotiff; profile=cloud-optimized'].
For COG / image assets, use stackstac or odc-stac to stack across items.

Tests

Added two regression tests against the existing simple_item fixture (its only assets are COG + thumbnail, mirroring NAIP's shape). All existing core tests stay green; test_to_xarray_reference_file was already failing on main in my env (unrecognized engine 'kerchunk') and is unrelated.

Fixes #68.

[ghostiee-11]

Hey @jsignell, please check into this!!