Align pyramid downscaling with `ome-zarr-py` (`nearest` for labels; tighten tests)

Summary

https://github.com/scverse/spatialdata/pull/1068 adds image-pyramid generation via ome-zarr-py internals as an alternative to multiscale-spatial-image. After https://github.com/ome/ome-zarr-py/pull/516 lands, we can further align and simplify our implementation.

Our pyramid tests currently permit sizable differences from label pyramids produced by multiscale-spatial-image due to mismatched downscaling:

• multiscale-spatial-image: “nearest” (for labels)
• spatialdata/ome-zarr-py: “resize” (order 0 to preserve label values)

Note: even after the ome-zarr-py PR merges, per-axis scale factors remain unsupported in ome-zarr-py. This should ideally be upstreamed. Until then, we will keep a custom downscaling function, leveraging the new ome-zarr-py APIs introduced by the PR.

Action items

• After https://github.com/ome/ome-zarr-py/pull/516 is merged:
  • Use “nearest” for labels and align image defaults accordingly.
  • Tighten test tolerances (reduce or eliminate fraction_non_equal for i > 0).
  • Adopt ome-zarr-py’s 3D downscaling APIs to minimize custom logic.
• Propose upstream support for per-axis scale factors in ome-zarr-py and simplify our code once available. Or choose to maintain this functionality in our code.

First verify that ome-zarr-py PR 516 has landed, then inspect the image-pyramid implementation added in spatialdata PR 1068 and the existing pyramid tests. Align label downscaling and image defaults with ome-zarr-py, adopt its 3D APIs where applicable, and tighten the tests by reducing or removing fraction_non_equal for levels above zero. Per-axis scale-factor support remains a separate upstream concern.