Add oxidation solver with coupled diffusion, deformation, and GPU acceleration

[FilipovicLado]

Adds a physics-based silicon oxidation model targeting LOCOS (Local Oxidation of Silicon) and fin-/step-/trench-oxidation geometries. The solver tightly couples oxidant diffusion, oxide viscous flow (Stokes), and nitride mask bending/creep into a single iterative loop.

New headers (include/viennals/):

• lsOxidation.hpp: top-level entry point; wires up the coupled solver

• lsOxidationDiffusion.hpp: FEM diffusion solver (Deal.II back-end)

• lsOxidationDeformation.hpp: Stokes flow for oxide viscous deformation

• lsOxidationMask.hpp: nitride mask bending and Arrhenius creep

• lsOxidationModel.hpp: SIMPLE-iteration coupling loop

• lsOxidationSolverBase.hpp: shared base for all sub-solvers

• lsOxidationPresets.hpp: material presets for common oxidation conditions

• lsOxidationBiCGSTAB.cuh / lsOxidationBiCGSTABInterface.hpp: GPU BiCGSTAB linear solver (CUDA Runtime; no OptiX dependency)

• GPU back-end (gpu/):

• lsOxidationBiCGSTABKernels.cu: CUDA kernels; enabled via VIENNALS_USE_GPU; falls back to CPU when unavailable

Physics highlights:

• Deal.II FEM for O2/H2O diffusion with concentration-dependent boundary conditions
• Stokes-equation viscous flow for oxide growth and volumetric expansion
• Two-way traction coupling at the mask-oxide contact (Neumann BC on mask, reaction force fed back to oxide)
• SIMPLE iteration for outer coupling (unconditionally stable)
• CFL-limited advection with cross-step warm starts for efficiency
• Dry and wet oxidation; Si⟨100⟩, ⟨110⟩, ⟨111⟩, and poly-Si orientation ratios
• Python bindings updated for all new classes; typed stubs included.

Tests added for OxidationModel (regression) and OxidationConvergence (grid-refinement order check).
Documentation: docs/OxidationSolver.md covers the governing equations and solver parameters.

[tobre1]

The LOCOS example doesn't seem to work out of the box for me. I get the following error thrown:

    LOCOS: starting time step, requested_dt=0.350000 hr
    DEBUG: LOCOS: solving coupled diffusion/deformation field for dt=0.350000 hr
    DEBUG: OxidationModel: coupling iteration 1/8 starting diffusion solve
    OxidationDiffusion: using CPU BiCGSTAB for diffusion solve (nodes=563, GPU mode not selected).
    diffusion n=563 iters=29 res=0.000336 [CPU] CPU BiCGSTAB took: 0.005837 s 
    diffusion n=563 [CPU] diag/b precompute took: 0.000102 s 
    DEBUG: OxidationModel: diffusion solve complete, nodes=563, iterations=29, residual=0.000336
    DEBUG: OxidationModel: coupling iteration 1/8 starting deformation solve
            mechanics[0] stokes   iters=41/100 res=0.000000 took: 0.068182 s 
        mechanics[0] pressure iters=34/500 res=0.000000 coupling=1.000000 took: 0.010382 s 
            mechanics[1] stokes   iters=50/100 res=0.000000 took: 0.079645 s 
        mechanics[1] pressure iters=35/500 res=0.000001 coupling=0.908977 took: 0.011208 s 
    WARNING: solveMechanics: did not converge after 2 iterations (residual=0.908977, tolerance=0.000000)
          deformation n=563 harmonic took: 0.061264 s 
      deformation n=563 mechanics-total took: 0.181244 s 
    DEBUG: OxidationModel: deformation solve complete, nodes=563, iterations=40, residual=0.908977
    WARNING: OxidationModel: deformation solve failed (mechanics=0.908977, pressure=0.000001, stokes=0.000000).
      coupled(iter=1) took: 0.261665 s 
    DEBUG: LOCOS: coupled diffusion/deformation solve complete
terminate called after throwing an instance of 'std::runtime_error'
  what():  LOCOS: coupled solve failed (deformation solve failed (mechanics=0.908977, pressure=0.000001, stokes=0.000000)).

[FilipovicLado]

The LOCOS example doesn't seem to work out of the box for me.

Sorry, missed updating the LOCOS example in ViennaLS because I was using the example in ViennaPS for testing and development. The example has been updated, along with python bindings and a python version of the example.

[tobre1]

I modified the parameter interface a bit, so you'll have to probably also adjust it in ViennaPS