Relax the torch-geometric minimum in the pytorch extra

[prajwal-tech07]

The pytorch extra pins torch-geometric>=2.5.3:

[project.optional-dependencies]
pytorch = [
    "torch>=2.3.0",
    "torch-geometric>=2.5.3",
]

That floor is higher than anything in the code needs, and it makes the extra unusable from projects pinned lower. neural-lam pins torch-geometric==2.3.1, so weather-model-graphs[pytorch] can't be installed there at all. It came up in mllam/neural-lam#596.

I don't think the pin was ever a compatibility decision. It's been >=2.5.3 since the initial commit in b68158f, back when 2.5.3 was the latest release, so it looks like it was just whatever was newest at the time rather than a considered floor.

The torch-geometric surface here is small. save/neural_lam/deprecated.py uses from_networkx and mentions pyg.data.Data in type hints, and save/base.py imports the package only to set HAS_PYG. Both have been around far longer than 2.5.3. The current save path, to_torch_tensors_on_disk, doesn't import torch-geometric at all, so the extra only matters for the deprecated to_pyg.

I checked that it actually runs. wmg 0.4.0 against torch-geometric 2.3.1:

wmg : 0.4.0
torch-geometric: 2.3.1
HAS_PYG : True
components : ['m2m', 'g2m', 'm2g']
to_pyg wrote : ['g2m_edge_index.pt', 'g2m_features.pt', 'g2m_node_features.pt', ...]
g2m_edge_index : (2, 592) | g2m_features: (592, 3)

All nine files written with the expected shapes. So I'd suggest dropping the minimum to torch-geometric>=2.3.1. That's the version I tested and the one neural-lam pins. Older versions may well work too, I just haven't checked. torch>=2.3.0 can stay as it is.

One thing that probably explains how this drifted: as far as I can tell the extra is never installed in CI. ci-tests.yml installs torch==1.8.1+cpu and then runs pdm install, which doesn't pull in optional groups, so HAS_PYG is False and both to_pyg tests return early. That would mean nothing has ever exercised this floor. Might be worth a follow-up to install the extra in CI so the minimum is actually tested, but happy to keep that separate from this.
Happy to put up the one-line change if that's useful.