MACE model download fails on CI

[ceriottm]

Describe the bug
We make examples of data representation using MACE features in the https://chemiscope.org docs.
Since a couple of days, the CI job building the docs fails because it cannot fetch the MACE-OFF model.
Haven't been able to figure out why so I'm hoping this is due to some changes here you can comment on.
Link to chemiscope issue: lab-cosmo/chemiscope#533

To Reproduce
Steps to reproduce the behavior:
Run the MACE-based example in the CI

Expected behavior
The example runs and builds the sphinx-gallery example using MACE

Screenshots
CI Error:

        ../../python/examples/7-explore-advanced.py failed leaving traceback:
    
        Traceback (most recent call last):
          File "/home/runner/work/chemiscope/chemiscope/python/examples/7-explore-advanced.py", line 125, in <module>
            cs = chemiscope.explore(qm9_structures, featurizer=mace_off_tsne, properties=properties)
          File "/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.13/site-packages/chemiscope/explore/__init__.py", line 200, in explore
            merged_properties["features"] = featurizer(structures, environments)
                                            ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^
          File "/home/runner/work/chemiscope/chemiscope/python/examples/7-explore-advanced.py", line 93, in mace_off_tsne
            calculator = mace_off(**descriptor_opt)
          File "/home/runner/work/chemiscope/chemiscope/.tox/docs/lib/python3.13/site-packages/mace/calculators/foundations_models.py", line 267, in mace_off
            raise RuntimeError("Model download failed and no local model found") from exc
        RuntimeError: Model download failed and no local model found

[ElliottKasoar]

I started hitting this recently too (stfc/janus-core#684).

For me, the error that MACE catches and hides was along the lines of urllib.error.HTTPError: HTTP Error 429: Too Many Requests, so I think it's due to stricter rate-limiting (e.g. see: https://github.com/orgs/community/discussions/159123).

The URLs vary quite a bit between the various models, but I think the MACE-OFF small and large are most affected, as it seems that blob links are limited more than raw links:

            urls = dict(
                small="https://github.com/ACEsuit/mace-off/blob/main/mace_off23/MACE-OFF23_small.model?raw=true",
                medium="https://github.com/ACEsuit/mace-off/raw/main/mace_off23/MACE-OFF23_medium.model?raw=true",
                large="https://github.com/ACEsuit/mace-off/blob/main/mace_off23/MACE-OFF23_large.model?raw=true",
            )

I don't know if there was a reason for this difference in the URLs, but it may be enough to just swap blob for raw?

[ilyes319]

Hey @ElliottKasoar could you make a PR swapping the urls?

[CompRhys]

Same issue has been seen in torchsim CI since ~ Mar 5th. I swapped blob->raw (in addition to other lint/fix changes) in #1378 and pointed my fork in TorchSim/torch-sim#497 to check that this was sufficient to resolve issue.