Enhancement: Local SpatialData object with remotely accessed pyramidal OME-ZARR image

WSIs can account for > 90% of a Spatial Transcriptomics experiment.
This can limit the number of datasets i can host and open on my laptop or workstation

Describe the solution you'd like
I'd like to leave the WSIs (saved as pyramidal ZARR files) on a S3 bucket or on a distant SSH server, and work with SpatialData "as if" the WSI was local.

Describe alternatives you've considered
I can open a multiscale image in napari, overlay a label layer (with coarser pixels) for an annotation tasks.
Zoom in and out is very smooth.
Below is the code for display the image and 2 screenshots of the usecase.

import numpy as np
import zarr
import dask.array as da
import napari

def open_pyramidal_zarr_img(uri, v, layer_name=None, reorder_axes=False):
    store = zarr.open(uri, mode="r")
    multiscale_data = [da.from_zarr(store[key]) for key in store.array_keys()]
    if reorder_axes:

        if reorder_axes == "CYX -> XYC":
            multiscale_data = [da.swapaxes(data, 0, -1) for data in multiscale_data]

        else:
            raise ValueError(f"reorder_axes must be False or 'CYX -> XYC', got {reorder_axes}")
    v.add_image(multiscale_data, name=layer_name, rgb=True)

(EDIT: scalebar is wrong, obviously 🙄)

Loading is fast (10-20s). Zooming in is smooth enough in my opinion for typical tasks.

It would be very nice to be able to do the same in the napari-spatialdata context.

What i tried

# open local sdata object
sdata = sd.read_zarr(local_path)

# add distant image
store = zarr.open(distant_uri, "r")
image = Image2DModel.parse(da.from_zarr(store["0"]), dims=('c', 'x', 'y'), c_coords=["r", "g", "b"])
sdata["full_res"] = image

# interactive display
interactive = Interactive(sdata)

It takes > 1 min, probably loading all image (then zooming is fast).
Each time layer visibility is toggled, it takes another minute

Start by comparing the provided open_pyramidal_zarr_img example with the attempted Image2DModel.parse and Interactive workflow, including sd.read_zarr and the remote URI handling. Done means a remote pyramidal image can be added to a local SpatialData object without loading the whole image, while visibility changes and zooming remain responsive.