Fix quadratic cost in adaptive mesh refinement's coarse/fine cell-map construction

[pbrubeck]

Summary

adaptive_parent_child_cell_maps resolved each fine cell's coarse parent with PETSc's DMLabelGetValue, called once per fine cell. DMLabelGetValue should resolve a point in O(1); instead it scans every stratum of the label linearly until it finds the one containing the queried point. set_adaptive_parent_label gives every coarse cell its own stratum value, so on an adaptively refined mesh the per-fine-cell loop was effectively O(nfine · ncoarse), not O(nfine). Profiling adaptive_multigrid.py under -log_view showed refine_marked_elements dominated not by PETSc's refine_sbr mesh transform itself, but by time attributed to this uninstrumented loop.

adaptive_parent_child_cell_maps now walks the label by stratum instead, with DMLabelGetStratumSize/DMLabelGetStratumIS, one call per coarse cell. That uses PETSc's O(1) value→stratum hash map and touches each fine cell exactly once, so the loop is O(nfine + ncoarse). Output is unchanged — cross-checked bit-for-bit against the original point-wise result on refined netgen meshes.

On the AdaptiveMeshHierarchy demo (first 10 Dörfler-marked refinement levels of an L-shaped netgen mesh), adaptive_parent_child_cell_maps's share of time inside refine_marked_elements drops from 59% to 8%, and refine_marked_elements's own total drops by 55%. The dominant cost left inside refine_marked_elements is now legitimately PETSc's own mesh-refinement transform.

Collateral changes

_adapt_marked_cells and refine_marked_elements (firedrake/adapt.py) gained PETSc.Log.Event wrappers around each of their previously-unattributed sub-steps (label marking, set_adaptive_parent_label, the adaptLabel transform, Mesh() construction, adaptive_parent_child_cell_maps, re-marking, Netgen re-curving). Without them, -log_view attributed all of this work to refine_marked_elements's own self-time rather than the call that actually spent it, which is what obscured this cost in the first place.

The adaptive_multigrid demo's recorded multigrid iteration counts and convergence plot predated #5215's DMPlexTransform-based mesh adaptivity; both are regenerated here against current refine_marked_elements output.

Tests

tests/firedrake/multigrid/test_adaptive_multigrid.py already exercises adaptive_parent_child_cell_maps across firedrake/netgen meshes and process counts; it passes unchanged at nprocs 1/2/4.

🤖 Generated with Claude Code

[pbrubeck]

Thanks. CI failures are known. Merging