feat: observation-informed interpolator

[OpheliaMiralles]

Description

Introduces the model from https://arxiv.org/pdf/2509.00017 + dummy tests.

What problem does this change solve?

Allows interpolator to be imbalanced, e.g. be inferred with past data (observations) and future (nwp forecasts or ML forecasts) known data.

What issue or task does this change relate to?

#209 is also in the paper even if it can also be applied to other models.

Additional notes

There was no other test for tasks or models so I created some, I don't know if I missed them or did not look at the right place?

As a contributor to the Anemoi framework, please ensure that your changes include unit tests, updates to any affected dependencies and documentation, and have been tested in a parallel setting (i.e., with multiple GPUs). As a reviewer, you are also responsible for verifying these aspects and requesting changes if they are not adequately addressed. For guidelines about those please refer to https://anemoi.readthedocs.io/en/latest/

By opening this pull request, I affirm that all authors agree to the Contributor License Agreement.
@copilot

[anaprietonem]

Some context for why I changed the label of this PR - It's not clear from the description/naming if this new task requires observations datasets and if that's not the case - if then we could still use gridded datasets and reproduce @OpheliaMiralles results. Based on that, I think as it's a new task it would be good to discuss this at ATS to decide if that's a task we want to support right now and what test maintenance would require.

I hope that's a bit more clear now - also by no means changing the label it's a blocker for review, if people find it interesting or have capacity please engage with this!

[OpheliaMiralles]

Some context for why I changed the label of this PR - It's not clear from the description/naming if this new task requires observations datasets and if that's not the case - if then we could still use gridded datasets and reproduce @OpheliaMiralles results. Based on that, I think as it's a new task it would be good to discuss this at ATS to decide if that's a task we want to support right now and what test maintenance would require.

I hope that's a bit more clear now - also by no means changing the label it's a blocker for review, if people find it interesting or have capacity please engage with this!

Thanks for the heads up. Indeed, results are reproducible with gridded datasets, which is what I used for the paper as multiple encoders and datasets were not available at the time. I still think combining it with a 1-hour emulator would be a nice and easy way of doing nowcasting with anemoi without having to integrate observations directly in the 6-hour forecaster, but I'll wait for some people to engage (if not I'll close it after e.g. 2 months otherwise the maintenance is on my side :)).