combine() does not normalise a raw module passed as array_package

[mwcraig]

In combine() (ccdproc/combiner.py:1055-1059) the requested array package is handled with

if array_package is not None:
    try:
        xp = array_api_compat.array_namespace(array_package)
    except TypeError:
        xp = array_package

array_api_compat.array_namespace() takes arrays, not modules, so for a module it always raises TypeError (module is not a supported array type — verified for numpy and dask.array) and the fallback hands the raw module through unchanged. The intent was clearly to normalise to the array-api-compat namespace; as written the normalisation never happens.

A raw module then reaches code that uses array-API-only features (xp.bool, the device= keyword in xp.asarray/xp.zeros, xp.astype, ...) and fails, e.g. combine(files, array_package=dask.array) dies with TypeError: from_array() got an unexpected keyword argument 'device' (on main before #976 it died one call later in clip_extrema with module 'dask.array' has no attribute 'argsort').

PR #976 fixed the same problem in Combiner.__init__ by normalising with

xp = array_api_compat.array_namespace(xp.asarray(0))

combine() should do the same so the two entry points agree on what array_package / xp may be, and the array_package docstring should say a plain module is accepted and converted.

Regression test: combine([...numpy CCDData...], array_package=np) and, ideally, array_package=dask.array on a FITS-file input so the CCDData.read branch that converts data/mask/uncertainty into xp is exercised.

Follow-up from the review of #976.