When a group has no explicit spatial:transform, derive a GDAL-order geotransform
from spatial:bbox plus the array's own spatial shape (+ optional
spatial:registration, default "pixel").

The derivation lives in the loader's transform-fallback chain, ahead of the
coordinate-array path, so it uses the array shape already in hand rather than a
separate, drift-prone spatial:shape attribute (bbox alone is sufficient).
geozarr.rs stays a pure attribute decoder that surfaces the bbox and
registration; the loader applies them.

A malformed or degenerate/inverted bbox is non-fatal: it warns and falls through
to the coordinate arrays (then identity), matching how a bad coordinate array is
handled. When the bbox path derives a transform and the group declares no CRS, a
geographic CRS is inferred from the spatial dim names, mirroring the
coordinate-array path.

Move the bbox-to-geotransform derivation out of sedona-raster-zarr into
AffineMatrix::from_bbox_and_spatial_shape (with a to_gdal_geotransform helper),
so it is reusable beyond Zarr. Add a rasterio from_bounds cross-check test. In
the Zarr loader, split the transform-resolution block into resolve_group_transform
/ transform_from_bbox / transform_from_coordinate_arrays.