Update NEMO ingestion code

[VeckoTheGecko]

Upon further testing, our convert.nemo_to_sgrid function needed some changes. This PR:

• Updates NEMO fieldset ingestion so that it expects depthw to be provided (and distinguishes from depths that are on the cell centers vs the cell edges) In fact, I guess they should be able to not provide depth in the 2D case. The convert module should be refactored later
• warns the user if they don't provide the depth

Still to add tests... Wondering what the best way to do that is for the convert.py module

xref Parcels-code/parcels-benchmarks#40

EDIT: Updated description to remove out of date test failure info

[VeckoTheGecko]

Also, another question about API:

Is the following the API we're going for?

    U = xr.open_mfdataset(data_folder.glob("*U.nc4"))
    V = xr.open_mfdataset(data_folder.glob("*V.nc4"))
    depth_da = xr.open_mfdataset(data_folder.glob("*W.nc4"))['depthw']
    coords = xr.open_dataset(data_folder / "mesh_mask.nc4")
    coords['depthw'] = depth_da
    ds = convert.nemo_to_sgrid(fields=dict(U=U, V=V), coords=coords)

and then if they want to do 2D advection (no depth)

    U = xr.open_mfdataset(data_folder.glob("*U.nc4"))
    V = xr.open_mfdataset(data_folder.glob("*V.nc4"))
    coords = xr.open_dataset(data_folder / "mesh_mask.nc4")
    ds = convert.nemo_to_sgrid(fields=dict(U=U, V=V), coords=coords)

(currently the test cases are failing since it expects depthw to be provided regardless - so explicitly providing [0] for 2D advection)

[erikvansebille]

Yep, I think that's a good API. Mirrors also what's now in Parcels v3, where 2D advection at the surface does not require a depth field (depth is then assumed zero throughout)

[fluidnumerics-joe]

@fluidnumerics-joe what's your experience working within parcels_benchmarks?

The MOI Curvilinear benchmark has been broken for some time. After getting the time ordering sorted out on my system, it's been a bit of whack-a-mole to get that benchmark running.

[VeckoTheGecko]

I'm going to park this PR for the timebeing to focus on other profiling stuff

The MOI Curvilinear benchmark has been broken for some time. After getting the time ordering sorted out on my system, it's been a bit of whack-a-mole to get that benchmark running.

Good to know - let me know when you'd like to hop back on that line, happy to work on it together

[VeckoTheGecko]

I'm going to merge this so that we can also move forward with Parcels-code/parcels-benchmarks#40 (I will then work on this "middle layer" of testing which is described in that PR so that we can better develop the convert module here).