wrap: true folds by layer, so the main chain stops reading in order #22

Split out of #18, whose "neighbouring problem" section this is. #21 fixes that issue's stranded arrowhead; this half is untouched by it.

What happens

wrap: true gives the best aspect ratio of anything we produce, but it folds on ELK's layer index, not on the chain. The main chain therefore stops being contiguous: a reader following it is sent backwards up the page.

Measured over the bundled examples and the agent-run diagrams, all forced to wrap: true — reading rank of each node on the longest forward chain, in reading order (row band, then x):

examples/enterprise.yaml rank=[6,3,4,5,8,19,10,11,12,13,14,15] 2 inversions
examples/growth.yaml rank=[5,1,2,3,6,9,21,12,13] 2 inversions
codex round 2 rank=[5,7,8,2,1,3,4,6] 2 inversions
agy round 2 rank=[2,3,6,7,8,0] 1 inversion — the chain ends at rank 0
examples/startup.yaml rank=[0,1,2,3] reads in order
examples/web-app.yaml rank=[0,1,4,5,6] reads in order

The last node of the agy round-2 diagram lands top-left. In an earlier round the pipeline's first node ended up bottom-right.

Why it cannot be done through ELK

Investigated while fixing #18:

• ELK cuts by layer index. Reordering nodes before handing them over does not move a cut.
• elk.layered.wrapping.cutting.strategy: MANUAL with cutting.cuts is the only option that places the folds explicitly, and it cannot be set from elkjs: the list type is not registered with ElkReflect, so the call throws Couldn't create new instance of property 'org.eclipse.elk.layered.wrapping.cutting.cuts'.
• elk.layered.considerModelOrder.strategy in all three modes (NODES_AND_EDGES, NODES, PREFER_EDGES) changed the output by not one pixel.

So the fold has to become ours: lay out once to get the chain order, partition it into rows, and emit each row for ELK to place — which is the "wrap becomes generated groups" shape.

Open question before starting

How a generated row interacts with a group the user wrote, when a group spans a row boundary. That has no answer yet, and it is the reason this is not a small change.

The detour warning added in 0.7.5 already names the fold as the cause, which is honest but does not make the picture readable.