Update anisotropic convection demo for current APIs

[ss2098]

This PR updates scripts/aniso_convection_demo.py so the annulus convection / metric-redistribution demo runs with the current Underworld3 APIs.

Changes include:

• convert AdvDiffusionSLCN estimate_dt() output to a scalar before passing it to solve()
• replace the unstable upper natural Stokes boundary condition with a stable upper essential velocity boundary for this demo
• create the cache directory before writing the cached convection state
• replace the retired smooth_mesh_interior(..., method="anisotropic") path with uw.meshing.node_redistribution()
• update the printed refinement label to match the current node_redistribution path

This does not change solver internals.

Validation:

• rm -f /tmp/metric_mesh/conv_ra1e+05_res16_n20.npz
• pixi run python scripts/aniso_convection_demo.py

Example local output:

• cached → /tmp/metric_mesh/conv_ra1e+05_res16_n20.npz
• === refine: node_redistribution on ρ∝|∇T| ===
• saved /tmp/metric_mesh/aniso_convection.png

[ss2098]

The failing test_uw3 import-smoke check appears to fail during MPI/UCX initialization before this PR change is reached:

UCX ERROR mana_0 ... failed to create UD QP
MPI_Init_thread failed
MPIDI_UCX_init_worker Input/output error

This PR only updates scripts/aniso_convection_demo.py. Local validation passed with:

rm -f /tmp/metric_mesh/conv_ra1e+05_res16_n20.npz
pixi run python scripts/aniso_convection_demo.py

The script completed and saved /tmp/metric_mesh/aniso_convection.png.

I do not have permission to rerun the upstream workflow, so a maintainer rerun may be needed.

[lmoresi]

Adversarial review

The os.makedirs fix for the /tmp cache is real, and the API updates verified correct (_tri_cells/_signed_areas still exported; node_redistribution(mesh, <expr>) valid). One blocking finding:

1. MAJOR — free-slip → no-slip on the upper boundary. This flips the demo's convection solution (rigid-lid vs free-slip annulus plumes differ visibly), and the demo exists to show refinement following those plumes. Use add_rotated_freeslip_bc(0, "Upper").
2. MINOR — from underworld3.meshing import smooth_mesh_interior is now unused; the helper block sits mid-imports (E402).
3. MINOR — zero_init_guess=True per step adds cost; harmless in a demo but unnecessary.

Requested change: the boundary condition. The failing test check was the known runner flake (#461), rerun.