erf.ibseb: store the checkpoint face state by face rather than as a cell field

[asalmgren]

Follow-up to the surface energy balance on immersed-boundary building faces (#3960), recording work deliberately deferred from that PR so it is not lost when the PR merges.

Where it stands

IBSEBState is a cell-centred MultiFab written once per level into the checkpoint. It started as a field over the whole level with 6 * (2 + n_slab_layers) components at every cell; #3960 shrank it twice, first onto column blocks around the buildings and then onto 4 x 4 blocks clipped to the k-range that owns faces, with one slot per face of a cell rather than a fixed six.

Measured on Exec/CanonicalTests/SEB/FaceStorage (a 128^3 level, one skyscraper, 2056 faces, n_slab_layers = 4):

layout	boxes	cells	components	values
whole level	-	2097152	36	75497472
8 x 8 blocks, ground to roof	8	15616	36	562176
4 x 4 blocks, clipped, n_slots slots	32	7872	18	141696
one record per face (ideal)	-	-	-	12336

So the field is still about 11.5x larger than the data it carries. The faces are a one-cell shell, and a cell field covering that shell with whole boxes cannot be proportional to the face count.

What would fix it

Store the state by face: one record of 2 + n_slab_layers reals keyed by (i, j, k, dir, side), nface records rather than O(cells covered).

Two routes, each with a genuine cost:

1. An AMReX particle container. Position at the fluid cell centre so each record lands in the grid that owns its face, dir/side as integer components. Checkpoint / Restart / Redistribute then give rank-count-independent restart for free, which is exactly the part Surface energy balance on immersed-boundary building faces (erf.ibseb), with immersed-boundary fixes for the forcing and the MRF/YSUNew schemes #3960 had to hand-roll with AllGatherBoxes plus a transfer MultiFab and a ParallelCopy. On restart, join the particles against the face list build() regenerates, per tile. The catch: ERF_USE_PARTICLES is an opt-in compile definition (CMake/BuildERFExe.cmake:199), so erf.ibseb would have to require particles or carry two checkpoint paths.
2. A sorted COO side file. Avoids the dependency, but means reimplementing the redistribute-onto-a-different-rank-count logic that the particle container already has, which is the part most likely to harbour a bug.

Also worth knowing

state_ncomp() is n_slots * (2 + n_layers()), where n_slots is the globally reduced largest face count on any one cell of the level. It is 1 for a convex isolated box but 3 on the FaceStorage skyscraper, because three faces meet on a cell at the rim. A single one-cell slot anywhere on a level pushes it to 6 and erases that part of the win. It is never worse than the fixed six the layout used to assume, but the saving is data-dependent in a way face-indexed storage would not be.

Relevant code: Source/ImmersedBoundarySEB/ERF_IBFaceSet.{H,cpp} (state_ncomp, state_boxarray, make_state, save_state, load_state) and Source/ImmersedBoundarySEB/ERF_IBSEB.cpp (init_ibseb, ibseb_write_checkpoint). Exec/CanonicalTests/SEB/FaceStorage covers the round trip, the restart onto another rank count, and the two mismatch aborts, so it should carry any replacement straight across.