Push derived variables beyond user interface

[darothen]

In the demonstration AR analysis script, we have to provide the following lines when we initializing an evaluation:

hres_forecast = inputs.ZarrForecast(
    source="gs://weatherbench2/datasets/hres/2016-2022-0012-1440x721.zarr",
    name="HRES",
    variables=[
        derived.AtmosphericRiverVariables(
            output_variables=["atmospheric_river_land_intersection"]
        )
    ],
    variable_mapping=inputs.HRES_metadata_variable_mapping,
)

Notably, we have to explicitly tell EWB that we want to derive a variable ("atmospheric_river_land_intersection"). But there is no way to know this a priori. Furthermore, we don't provide any specification to EWB to tell it how to compute this variable, beyond assuming that it will capture any naming harmonization with the "variable_mapping" argument.

The design here should be re-factored to push this sort of derived variable down the stack. It's a consequence of the event types of the cases we provide - not of the input dataset itself - and should logically be handled by that area of the library.

[linear]

EWB-81

[aaTman]

I am convinced at this point that DerivedVariables is not necessary and we can simplify this aspect of the library, simultaneously handling the issues you bring up here.

The primary challenges (in my eyes) will be:

1. How to provide said a priori approach to continue the philosophy of a simple approach to assigning a derived variable. We can definitely simplify the API here, just need to think intentionally on it
2. Avoiding pitfalls in computation e.g. ARCO-ERA5's None chunks preventing it being included in preprocessing - considering the importance of ERA5 as a target dataset I am still very hesitant to slow the code down by swapping the chunks default away from None
3. Building this to slot into the existing pipeline as a (more or less) drop in replacement for DerivedVariable

[darothen]

One idea I had was to just plug into the existing infra for "variable_mapping". A better approach which I nearly have a demo for is actually to build factories that create the actual event evaluation objects. Will have something to share today or tomorrow.