[feature] Stretched graphs

[JPXKQX]

This PR implements the functionality to build stretched graphs only in an area of interest (AOI).

This includes:

• The aoi_mask_builder objects is renamed to area_mask_builder.
• A new node builder, StretchedTriNodes, which allows to create a stretched graph.
• Expanding the MultiScaleEdges functionalities to support multi-scale connections in stretched graphs.

Checklist before requesting a review

• I have performed a self-review of my code
• My code follows the style guidelines of this project
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation and docstrings to reflect the changes
• I have added tests that prove my fix is effective or that my feature works
• I have ensured that the code is still pip-installable after the changes and runs
• I have ran this on single GPU
• I have ran this on multi-GPU or multi-node

Issue ticket number and link

Issues #8

This work would not have been possible without @theissenhelen and @JesperDramsch and the rest of the AIFS team! Thanks also to Magnus, Håvard and the MetNorway team for their contributions to this PR.

---

📚 Documentation preview 📚: https://anemoi-graphs--51.org.readthedocs.build/en/51/

[havardhhaugen]

Overall this looks really good, and we're pretty much able to reproduce the stretched grid graphs from the old graphgen in aifs-mono. I do have two suggestions:

• Add a keyword argument min_distance_km to CutOutZarrDatasetNodes so this can be set in the config (Or add functionality for passing any kwargs to CutOutZarrDatasetNodes through the config).
• Change how edge_length.invert works:

1. Currently the edge length normalization looks weird when used with invert; in the encoder / decoder I get edgelengths between [0.96,1] and in the processor some of them have negative edge weight. I'm guessing this happens because edgelengths are inverted before the call to super().post_process(values) in EdgeLenght.post_process().
2. Change the default value of invert to be True, giving shorter edges higher weight by default.

[JPXKQX]

Thanks for taking the time to review the PR.

• The CutOutZarrDatasetNodes will be refactored to receive a dict-like argument dataset directly from the configuration file. This change will be merged in feat: support area masking of boundary forcing #52 .
• We left invert=False as the default because we didn't see any difference (in performance). Regarding the difference between encoder/decoder and processor edge lengths, do you use the same norm in both cases? Would you expect to invert after normalisation?

[havardhhaugen]

I used unit-max normalization for both enc/dec and proc. The way invert is implemented (length = 1 - length), I would expect this to be done after the edge lengths are normalized to [0,1].

[JPXKQX]

The issue should now be resolved. Invert is applied after normalization, and a new method called "unit-range" has been introduced to map values to the [0, 1] interval, as "unit-max" only scales by dividing by the maximum.

[havardhhaugen]

Great! The edge lengths with invert and unit max now look like what I would expect.

[mchantry]

LGTM, thanks very much.