[Feat] Support irregular/unstructured grid graphs via Delaunay triangulation

[jainsaksham2006]

Motivation

Currently weather-model-graphs only supports rectilinear mesh layouts.
This limits neural-lam to datasets on regular grids, excluding:

• Unstructured NWP model output (e.g. ICON icosahedral, OpenIFS)
• Sparse observational data (ship tracks, surface stations, radiosondes)

Proposed Solution

Introduce mesh_layout="delaunay" as a new archetype that:

1. Accepts an arbitrary set of (lat, lon) coordinates as input
2. Computes a Delaunay triangulation via scipy.spatial.Delaunay
3. Optionally applies Lloyd relaxation for mesh quality improvement
4. Builds g2m, m2m, m2g edges using KD-tree neighbour search
5. Returns output in the same NetworkX format as existing archetypes

Graph Quality Metrics

Alongside this, I propose adding a wmg.metrics.graph_quality_report()
function that computes:

• Node degree variance
• Edge length coefficient of variation
• Triangle quality (min angle distribution)
• Coverage ratio

Questions for maintainers

• Is mesh_layout="delaunay" the right API extension point, or would
  you prefer a separate factory function?
• Should sparse observational data (different node type from mesh nodes)
  be handled in the same PR or a follow-up?

This is part of my GSoC 2026 proposal for Project 1 (Flexible Graph Construction).

— Saksham Jain (@jainsaksham2006)