Synoptic obs extraction doesn't account for model coordinate cutouts

### Describe the bug

Originally reported by Kirsty:

> Basically, I'm using a cutout subarea which I originally defined in model relative coordinates but as I had USE_WMO_STATION_NUMBERS=False the fetch_obs task was trying to retrieve all the MetDB obs within my cutout subarea but failing as it was looking for obs near the equator! Once I changed SUBAREA_EXTENT to be in real world coords, the fetch_obs task worked. So it would be useful if the Help panel could be updated to say that SUBAREA_EXTENT needs to be in real world coords if comparing with synoptic obs.

### How to reproduce

Steps to reproduce the behaviour:

1. Run CSET with synoptic observations enabled, USE_WMO_STATION_NUMBERS=False, and a cutout of a rotated pole model done using SUBAREA_TYPE=model.
2. Observe the model crashes due to not finding any observations, as it looks for them in the wrong place.

### Expected behaviour

Even when using model coordinates to cut out a domain, the correct observations should be fetched.

Start with the fetch_obs task and the SUBAREA_EXTENT/SUBAREA_TYPE=model handling described in the reproduction steps. Reproduce the rotated-pole cutout with USE_WMO_STATION_NUMBERS=False and compare model-coordinate and real-world-coordinate behavior. Done means synoptic observations are fetched from the correct geographic area for model-coordinate cutouts, and the Help panel states the coordinate requirement if that remains part of the fix.

Written by the indexing model from the issue text.