cavity: reset restart hnode to the mesh nominal in restart_thickness_ale (moving cavity)

[JanStreffing]

What

Fixes a persistent tracer-content drain when restarting a run whose ice-shelf cavity geometry has changed between legs — i.e. a moving-cavity coupled setup (FESOM ↔ an ice-sheet model such as PISM, where the ice-shelf draft is updated and the mesh regenerated/remapped each coupling leg).

Why it only bites a moving cavity

In restart_thickness_ale, cavity columns (and their open-ocean edge ring — any node touching a cavity element, ulevels_nod2D_max > 1) keep their layer thicknesses fixed at the mesh nominal (hnode_new from init_thickness_ale, the same exclusion class as the per-step hnode↔hnode_new sync loop). The previous code cycled these columns on the restart read.

With a static mesh that is harmless: the restart hnode equals the nominal, so nothing drifts. But when the cavity moves, a column that was open-ocean zstar in the leg that wrote the restart can be cavity (or edge-ring) in the leg that reads it, so its remapped restart hnode no longer matches the new mesh's nominal. Because the column is cycled, that mismatch persists forever, and the tracer T* update then applies it as a multiplicative content drain every step.

Fix

For cavity/edge-ring columns, force restart hnode to the nominal (hnode_new) and rebuild the depth arrays (zbar_3d_n, Z_3d_n) before cycle. Open-ocean columns are unchanged, so static-mesh and non-cavity runs are entirely unaffected.

[JanStreffing]

Marking as draft — not merge-ready yet.

In practice this hnode-to-nominal reset only partly addresses the moving-cavity problem under zstar. It removes the multiplicative tracer-content drain from a remapped restart, but it does not recover cavity velocities to the quality we get under linfs: zstar moving-cavity columns still show elevated, persistent barotropic currents relative to a linfs run on the same mesh (the excess appears injected across the mesh increment and is cavity-wide and persistent). So the fix is necessary but not sufficient.

Reaching linfs-like velocities under zstar needs further work on the momentum/thickness handling across a mesh increment, not just the restart hnode seeding. Keeping this open as a partial correctness fix; it should stay in draft until that gap is closed.

(Note: it is inert for linfs — the whole block is guarded by which_ale=='zstar' .or. 'zlevel', and the linfs restart path rebuilds hnode from the reference depth levels regardless.)