Refactor flat_multiscale connectivity to support crop_to_convex_hull

[AdMub]

Description

Following the implementation of spatial mesh pruning via Delaunay triangulation in PR #58, the crop_to_convex_hull feature currently supports keisler (flat) and hierarchical architectures.

However, it currently raises a ValueError if used with flat_multiscale (GraphCast) graphs.

The Bottleneck

The limitation exists because create_flat_multiscale_from_coordinates (in mesh/connectivity/flat.py) merges hierarchy levels by reshaping the node array under the assumption of a full (nx, ny, 2) rectangular grid and striding over indices. Once nodes are pruned spatially, this array reshaping fails.

Proposed Solution

To fully support multiscale graphs, the merging logic needs to be refactored to be position-based (spatial) rather than index-based. This will allow the graph to correctly connect coarser levels to the finest levels even when the grid is irregular or spatially cropped.

Note: This issue was spun off from PR #58 to prevent scope creep on the v0.4.0 release.

[AdMub]

Hi @leifdenby, that is a massive compliment, thank you! I completely agree that we should highlight it so it doesn't get missed.

1. I just opened the tracking issue for the flat_multiscale refactor here: Refactor flat_multiscale connectivity to support crop_to_convex_hull #145.
2. I think adding a subsection to docs/creating_the_graph.ipynb makes the most sense, since this is fundamentally a layout/creation constraint. I already have the matplotlib code written from my local testing that plots a side-by-side "Before" and "After" of a spatially cropped dataset.

I have added that notebook subsection, run the formatting, and push it up today!