add subsetting also in data loaders, not just in downlaoder

[JoachimKoenigslieb]

Fixes an issue where subsetting is only perfomed in the download branch of run_mode. This means the running on local files is very slow since we do the optical flow on the whole satellite domain!

To achieve this I have:

• moved subset_to_bbox to geospatial to avoid circular imports
• call subset_to_bbox in all branches of run_mode in main

[KristianHMoller]

Unfortunately, in the current form, this does not work with out MSG-CPP data from KNMI, as it results in an empty y-dimension:

<xarray.Dataset> Size: 6kB
Dimensions:  (time: 4, y: 0, x: 713)
Coordinates:
  * time     (time) datetime64[us] 32B 2026-05-06T09:15:00 ... 2026-05-06T10:...
  * y        (y) float64 0B 
  * x        (x) float64 6kB -10.73 -10.69 -10.64 -10.6 ... 19.89 19.94 19.98
Data variables:
    crs      |S1 1B b''
    sds      (time, y, x) float32 0B 
    sds_cs   (time, y, x) float32 0B 

I will do a bit more digging and get back to you.

[KristianHMoller]

Okay, this is a little odd, but the reason it fails is that for the MSG-CPP dataset (at least the way we download it), the latitudes are inverted. So the slice cuts out all data.

<xarray.Dataset> Size: 2MB
Dimensions:  (time: 1, y: 368, x: 713)
Coordinates:
  * time     (time) datetime64[us] 8B 2026-05-05T15:15:00
  * y        (y) float64 3kB 63.48 63.43 63.39 63.35 ... 47.4 47.36 47.32 47.27
  * x        (x) float64 6kB -10.73 -10.69 -10.64 -10.6 ... 19.89 19.94 19.98

So perhaps subset_to_bbox could end similar to this rather than our existing implementation which seems to work only for DWD data.

    # Handle potentially inverted coordinates by detecting order
    lat_values = ds[lat_coord].values
    lon_values = ds[lon_coord].values
    
    lat_ascending = lat_values[0] < lat_values[-1]
    lon_ascending = lon_values[0] < lon_values[-1]
    
    # Arrange slice bounds based on coordinate order
    lat_slice = slice(lat_min, lat_max) if lat_ascending else slice(lat_max, lat_min)
    lon_slice = slice(lon_min, lon_max) if lon_ascending else slice(lon_max, lon_min)

    return ds.sel(
        {
            lat_coord: lat_slice,
            lon_coord: lon_slice,
        }
    )

[KristianHMoller]

I think this is a very elegant solution to the issue of not cropping the data in the cases of files and s3!

[JoachimKoenigslieb]

Have added your suggested code to also handle both "directions" of latlon grids.

Also added changelog and ran the pre-commit hook!