feat: add variable-resolution latent graph builder

[AswaniSahoo]

Pull Request

Description

Builds on #226. build_variable_resolution_mesh returns a mixed-resolution cell set, coarse over the globe and fine over the region, and this PR turns that set into the latent graph the processor message-passes over.

The problem is the seam between the two resolutions. Same-resolution neighbours come from grid_disk, but that never crosses from a coarse cell to a fine one, so on its own the fine region is a disconnected island and no information flows in or out. Each fine cell is therefore also linked to the coarse cells bordering its coarse ancestor, found with cell_to_parent and then grid_disk on that parent, keeping the coarse cells that are actually present in the mesh. Edges are bidirectional and carry [sin d, cos d] of the great-circle distance between cell centres, the same edge feature the single-resolution builder already uses.

Pure function: it takes a cell list and returns a torch_geometric Data. No model or training changes yet - wiring it into the forecaster and resizing the embedding table come in later PRs.

Related to #3

How Has This Been Tested?

Added tests/test_stretched_latent_graph.py - plain pytest, runs entirely on h3, no network:

• at least one edge joins a coarse cell to a fine cell, so the seam is bridged
• union-find over every edge gives a single connected component, so the fine region is not an island
• two adjacent fine interior cells still share an edge (same-resolution case intact)
• no self-loops
• no duplicate edges
• edge indices reference exactly the supplied cells
• edge_attr is [E, 2] with each row on the unit circle

Commands:

• pytest tests/test_stretched_latent_graph.py -q -> 7 passed
• ruff check graph_weather/models/layers/stretched_latent_graph.py tests/test_stretched_latent_graph.py -> clean

Checklist:

• My code follows OCF's coding style guidelines
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• I have checked my code and corrected any misspellings