Load flat graphs into pyg.HeteroData and wire into BaseGraphModel

[prajwal-tech07]

First step of the HeteroData migration (#385), scoped to flat (non-hierarchical) graphs.

• Adds neural_lam.utils.heterodata with graph_dict_to_heterodata and its inverse heterodata_to_graph_dict. Grid nodes are the grid node type and mesh nodes the mesh node type, matching the terminology used throughout neural-lam (grid_static_features, mesh_static_features, g2m/m2m/m2g); g2m/m2m/m2g become typed edges. The grid node count is taken from the datastore rather than inferred from edge indices, and the edge-feature width is read dynamically (3 for 2D, 4 for 3D). The node/edge-type names are module-level constants; the reference implementations linked from the issue (leifdenby/weatherduck, matschreiner/equicast) use data/hidden - happy to switch if preferred.
• Wires into BaseGraphModel behind a use_heterodata flag (default False): when enabled the graph is represented as a HeteroData object and the model's graph tensors are taken from it. The extracted tensors are identical to the dict ones, so the model builds and trains identically either way, and existing checkpoints remain compatible.
• Threads the flag through GraphLAM.

Hierarchical (HiLAM) support is deliberately left as a follow-up - there is an open question on #385 about how to represent mesh levels as node/edge types.

Issue Link

Part of #385 (first implementation step; the design discussion and hierarchical follow-up remain open).

Type of change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📖 Documentation (Addition or improvements to documentation)

Checklist before requesting a review

• My branch is up-to-date with the target branch - if not update your fork with the changes from the target branch (use pull with --rebase option if possible).
• I have performed a self-review of my code
• For any new/modified functions/classes I have added docstrings that clearly describe its purpose, expected inputs and returned values
• I have placed in-line comments to clarify the intent of any hard-to-understand passages of my code
• I have updated the README to cover introduced code changes
• I have added tests that prove my fix is effective or that my feature works
• I have given the PR a name that clearly describes the change, written in imperative form (context).
• I have requested a reviewer and an assignee (assignee is responsible for merging). This applies only if you have write access to the repo, otherwise feel free to tag a maintainer to add a reviewer and assignee.

Checklist for reviewers

Each PR comes with its own improvements and flaws. The reviewer should check the following:

• the code is readable
• the code is well tested
• the code is documented (including return types and parameters)
• the code is easy to maintain

Author checklist after completed review

• I have added a line to the CHANGELOG describing this change, in a section
  reflecting type of change (add section where missing):
  • added: when you have added new functionality
  • changed: when default behaviour of the code has been changed
  • fixes: when your contribution fixes a bug
  • maintenance: when your contribution is relates to repo maintenance, e.g. CI/CD or documentation

Checklist for assignee

• PR is up to date with the base branch
• the tests pass
• (if the PR is not just maintenance/bugfix) the PR is assigned to the next milestone. If it is not, propose it for a future milestone.
• author has added an entry to the changelog (and designated the change as added, changed, fixed or maintenance)
• Once the PR is ready to be merged, squash commits and merge the PR.

[prajwal-tech07]

@leifdenby @joeloskarsson could I get a review on this when you have a chance? It's been rebased onto main after #648 - that PR moved graph loading into utils.load_and_register_graph, so the use_heterodata wiring now lives there and applies to any module loading a graph through that helper, including the new Graph-EFM predictors. CI is green and it's mergeable. #713 (hierarchical) stacks on top of this one.

[Sir-Sloth-The-Lazy]

HeteroData is neither Tensor, Parameter, nor Module, so this setattr lands in module.__dict__ and nn.Module._apply skips it. After .cuda(), the registered buffers move but module.graph keeps the original CPU tensors, a stale second copy of m2m_features etc. held on host memory. Harmless today since nothing reads module.graph back, but it'll bite the moment a later PR actually uses it in forward.

[Sir-Sloth-The-Lazy]

I think this is an overstatement, if you would like to clarify this ? I am sorry, I did not understand this

[Sir-Sloth-The-Lazy]

use_heterodata is unreachable outside a direct GraphLAM(...) call.

• BaseHiGraphModel.__init__ doesn't accept it and doesn't forward it to super().__init__(), so HiLAM/HiLAMParallel can never set it (moot anyway since they're hierarchical, but the flag isn't even reachable to raise the informative error).
• BaseGraphEFM.__init__ calls load_and_register_graph without use_heterodata at all.
• train_model.py defines no --use_heterodata CLI flag, so it can't be set from the training entrypoint either.

The PR thread says "the use_heterodata wiring now lives [in load_and_register_graph] and applies to any module loading a graph through that helper, including the new Graph-EFM predictors" that's true of the helper's signature, but no caller actually passes it through, so in practice the only way to exercise this flag today is constructing GraphLAM directly (which is what the test does).