Add FieldSet.from_sgrid_conventions()

[VeckoTheGecko]

Towards #2003

Changes:

• See commits and descriptions

The testing as it stands for from_conventions is just testing init on the datasets provided.

I think there's room for future improvements (which I've put as comments in the body)

This acts as the base for future methods (from_nemo etc) for ingesting structured data. Here is what I have in mind:

def from_nemo(field_data: dict[str, xr.Dataset], ds_mesh: xr.Dataset) -> xr.Dataset:
    # Not 100% sure about the function namings, location in the package, and the parameter API. I think that a single xr.Dataset should be returned so users can make modifications.

    ...  # work with field_data and ds_mesh
    ds = ...

    ds["grid"] = xr.DataArray(
        data=np.array([0]),
        attrs=sgrid.Grid2DMetadata(...).to_attrs(),
    )
    # ? does the "location" attribute need to be added to each field? (as per SGRID conventions)
    # ? would it be nice to use `sgrid.rename_dims()` to rename all dimensions to a common naming scheme used for Parcels from_* methods?
    # ? attach other cf-convention attributes? (standard_name etc.)
    # ? Add logger.info or logger.warning messages during ingestion for items the user needs to be aware of?

    return ds

# user code
ds_sgrid_compliant = from_nemo(...)

print(ds_sgrid_compliant)
# Looks good!
FieldSet.from_sgrid_conventions(ds_sgrid_compliant)
# OR
# Looks like the x variable is defined on the cell centers
ds_sgrid_compliant['x'].attrs['location'] = ...
# fixed
FieldSet.from_sgrid_conventions(ds_sgrid_compliant)

[erikvansebille]

Looks good! A few comments/questions below

[erikvansebille]

Why is it necessary to make a copy?

[VeckoTheGecko]

This is a shallow copy - from_fesom2 and from_copernicusmarine do them both to avoid changes to the dataset propogating to the original dataset

[erikvansebille]

But this interpolator may not be the default for nemo; so how would from_nemo overwrite the the interpolator here?

[VeckoTheGecko]

I think it should be the default no? Since XLinear would be aware of the grid positioning?

Otherwise we would need to expose this as a param to the function

[erikvansebille]

No, XLinear is only the default for an A-grid layout. For a C-Grid layout, the velocities should be CGrid_Velocity and any other field should be CGrid_Tracer. SO it depends on the Grid topology

[VeckoTheGecko]

Discussed in person. Outcome: I think we can find where the variables are on the grid via the sgrid conventions, and then use that to choose the interpolator. We'll merge for now, I can rework from_copernicusmarine to use this new code path, and we can explore in future how we can specify this in SGRID and adapt our interpolators