Multigrid linear prolongation with tigher full multigrid turbulence integration

[bigfooted]

Proposed Changes

Multigrid on stretched meshes was slow or unstable in three situations. The coarse grids differed with the number of MPI ranks. The coarse-grid correction was only piecewise constant. The CFL adaptation could not tell a stuck run from a converging one. This branch addresses these. Every new option is off by default.

Main changes

• Linear prolongation (MG_LINEAR_PROLONGATION). The coarse-grid correction is prolonged with a limited least-squares gradient over each coarse control volume, instead of injecting the parent value into every child.
  The reconstruction is exact for a linear correction, and the children still average back to the coarse value.
  Walls and symmetry planes keep the constant operator.
• Implicit-line agglomeration fixes.
  Narrow 1×2 patches are handled in the line paving.
  The agglomeration of RAE2822-type meshes is improved.
  Columns that reach a partition interface are handed to the neighbouring rank. The coarse-grid halo coordinates are taken from the owning rank, so coarse stencils don't depend on the partitioning.
• Coarse-grid viscous fluxes. The corrected gradient is now used on all levels, not only the fine grid. The viscous Jacobian uses the effective distance |d|²/|d·n|, which matches the corrected normal derivative.
• Full-MG startup with turbulence.
  Wall distance and roughness are restricted to the coarse grids and exchanged to halo points. The nearest-neighbour wall distance is recomputed on the coarse levels.
  Coarse-level residuals are reduced across ranks. MGLEVEL is checked against the supported maximum.
  The startup stagnation test no longer counts a growing residual as stagnation.
• some [split off as separate PR, independent of multigrid] BCGSTAB returns its best iterate. BCGSTAB does not reduce the residual monotonically, so the solver now returns the best iterate it saw instead of the last one. text
• Rank-independent partitioning (PARMETIS_ANISOTROPY_WEIGHT).
  Graph edges are weighted by the aspect ratio of their cells, and anisotropic lines are moved onto a single rank.
  Line edges are now identified by a rule that is stable under ties: an edge joins a line if it is the strongest at one end, or at least half the strongest at both ends. Before, a single "strongest edge" link per point fragmented lines on near-uniform spacing, and that is what made convergence depend on the rank count.
• Residual-driven CFL adaptation (CFL_ADAPT_RESIDUAL). The CFL follows the finest-grid nonlinear residuals, so it does not depend on how many ranks split the preconditioner.
  Growth: the CFL grows while the residuals fall.
  Jumps: a median jump across the flow equations, or a failed linear solve, halves the CFL. A spike of more than one decade in any single flow residual within 5 iterations does the same. Repeated events at similar CFL set a lasting cap.
  Stagnation: no progress for 150 iterations caps the CFL below its current value.
  Frozen residuals [independent of multigrid]: if the mean log flow residual, with turbulence left out, varies by less than 0.01 decades over 30 iterations, the CFL cap becomes 0.6 times the current CFL.
• [independent of multigrid] First-order reconstruction at wall ends (MUSCL_FIRST_ORDER_WALL_ENDS). MUSCL is dropped to first order on the edges of points where a no-slip wall ends on a surface-continuing boundary, or at a sharp edge. The gradient is singular there, and MUSCL across it can stall or destabilise multigrid.
  AD build fixes for the discrete-adjoint build.

PR Checklist

Put an X by all that apply. You can fill this out after submitting the PR. If you have any questions, don't hesitate to ask! We want to help. These are a guide for you to know what the reviewers will be looking for in your contribution.

• I am submitting my contribution to the develop branch.
• My contribution generates no new compiler warnings (try with --warnlevel=3 when using meson).
• My contribution is commented and consistent with SU2 style (https://su2code.github.io/docs_v7/Style-Guide/).
• I used the pre-commit hook to prevent dirty commits and used pre-commit run --all to format old commits.
• I have added a test case that demonstrates my contribution, if necessary.
• I have updated appropriate documentation (Tutorials, Docs Page, config_template.cpp), if necessary.

[bigfooted]

Main thing to go through is the adaptive CFL. It needs more attention anyway, because it is the best convergence accelerator that we have. For now this seems to work well on the cases I've tested.