array-api-strict: subtract_overscan and cosmicray_lacosmic(gain_apply=True) fail on integer data #1034

On array-api-strict, two ccdproc functions fail on integer data (uint16, int32 and so on). array-api-strict refuses the implicit integer → floating promotion that NumPy, dask and JAX do. The cause in both cases is ccdproc's own arithmetic, not a host round trip.

1. subtract_overscan, with or without model=:
   • median=False calls xp.mean on the integer overscan: TypeError: Only floating-point dtypes are allowed in mean.
   • median=True gets a float64 median, then fails at ccd.data - oscan: TypeError: array_api_strict.uint16 and array_api_strict.float64 cannot be type promoted together.
2. cosmicray_lacosmic with gain_apply=True, on an integer CCDData that has an uncertainty. The uncertainty is scaled with _ccd.multiply(gain_value), which multiplies the integer data by a floating 0-d array as well: TypeError: array_api_strict.uint16 and array_api_strict.float32 cannot be type promoted together (float64 in place of float32 on main before Keep the floating dtype of the input in the host round trips #1032).

Reproducer, run on array-api-strict device1 (the same happens on the default device):

import array_api_strict as xp
import numpy as np
import astropy.units as u
from astropy.nddata import CCDData, StdDevUncertainty
from ccdproc import subtract_overscan, cosmicray_lacosmic

dev = xp.Device("device1")
data = xp.asarray(
    np.random.default_rng(0).normal(1000, 10, (40, 40)).astype("uint16"), device=dev
)

for median in (False, True):
    subtract_overscan(
        CCDData(data, unit="adu"), overscan_axis=1, fits_section="[1:5,:]", median=median
    )  # TypeError

ccd = CCDData(
    data,
    unit="adu",
    uncertainty=StdDevUncertainty(xp.full((40, 40), 10.0, dtype=xp.float32, device=dev)),
)
cosmicray_lacosmic(ccd, gain=2 * u.electron / u.adu, gain_apply=True)  # TypeError

History: found while working on #1023 (PR #1032), and present on main. PR #1032 pins the subtract_overscan mean path and the cosmicray_lacosmic case with non-strict backend_xfail markers; the subtract_overscan median path has no test. #1013 (subtract_dark(scale=True) promoting float32 on NumPy) is a related dtype problem from the same kind of 0-d float64 operand, but it is a different bug. Summary noted on #971.