Add nls::petsc::SNESSolver, a C++ wrapper for PETSc SNES

[jhale]

Wraps PETSc SNES in dolfinx::nls::petsc::SNESSolver, adapting C++ callables to the SNES callback interface. It is intended to supersede the deprecated nls::petsc::NewtonSolver.

The class owns the SNES object and holds references to the vector and matrices that define the layout of the objects the callbacks assemble into. Configuration of the solve is left to the user, via the options database or the SNES object returned by snes(). Lifecycle follows la::petsc::KrylovSolver: an MPI_Comm constructor, a (SNES, bool inc_ref_count) wrapping constructor, deleted copies, and noexcept moves.

nls::petsc::SNESSolver solver(mesh.comm());
solver.set_F([&](const Vec x, Vec b) { ... }, b);
solver.set_J([&](const Vec x, Mat A, Mat) { ... }, A);
solver.set_options_prefix("my_problem_");
solver.set_from_options();
solver.solve(x);

The name deliberately differs from the Python dolfinx.fem.petsc.NonlinearProblem. That class takes UFL forms, u and bcs, and creates its own matrices, vectors and SNES; this one takes callables and objects the caller owns. NonlinearProblem stays free for a form-based C++ class, should one be added.

Also included:

• fem::petsc::assemble_residual and fem::petsc::assemble_jacobian, C++ counterparts of the functions of the same name in dolfinx.fem.petsc, for use as the bodies of the callbacks. Their semantics were checked step by step against the Python versions: ghost updates, zeroing, lifting with alpha = -1, set_bc, and the guard that inserts a unit diagonal only for square forms.
• The C++ hyperelasticity demo uses both, replacing its hand-written Newton loop and its HyperElasticProblem class. It gains a line search, and converges in three Newton iterations rather than four because the hand-written residual did not lift the Dirichlet conditions.
• set_update, exposing SNESSetUpdate. That PETSc callback takes no context argument, so the solver is recovered from a PetscContainer composed on the SNES.
• Exceptions thrown by a callback are stored and re-thrown by solve rather than being reduced to a PETSc error code. PETSc does not restore its state as the aborted solve unwinds (x is left locked for read-only access), so this is documented as terminal for the solver.
• KrylovSolver::solve and SNESSolver::solve mutate solver state, so neither is const. KrylovSolver::solve losing const is the only change visible to existing code; the two call sites in the tree are unaffected.
• Fix for undefined behaviour when reading an unset options prefix: PETSc reports it as a null pointer, which was passed straight to the std::string constructor. Affected get_options_prefix on la::petsc::Vector, la::petsc::Matrix and la::petsc::KrylovSolver as well.
• C-style casts of PETSc handles in la/petsc.cpp replaced with reinterpret_cast.

Tested with cpp/test/nls/snessolver.cpp on one, two and three ranks: solve, solve after a move, wrapping an existing SNES, a separate preconditioner matrix, the update hook, exceptions from each of the three callbacks, non-convergence, the initial guess being respected, repeated solves, and PETSc reference counts before and after destruction.

The solver only ever handles Vec and Mat, so MATNEST and VECNEST need no support and pass through untouched. There is a test for this: two decoupled blocks with different roots, MPIAIJ under MPI, preconditioned with fieldsplit since a nest matrix cannot be factored directly.

Known gaps, for reviewers:

• The two new fem::petsc free functions have no unit test of their own; the hyperelasticity demo is their only coverage.
• The fem::petsc assembly functions handle single forms only. Blocked and nest assembly, which the Python class supports, are not implemented: they are built on VecGhostGetLocalForm, which does not apply to a VECNEST. This is assembly-helper work, not solver work.
• No Python bindings: Python users have petsc4py.
• assemble_jacobian takes the preconditioner form P as a trailing defaulted pointer, so it is separated from Pmat in the argument list. Passing nullptr in the middle of the list breaks template argument deduction.
• A SNES that outlives its SNESSolver carries callbacks pointing at a destroyed object. This is documented on the wrapping constructor rather than defended against: the function and Jacobian contexts cannot be cleared, as SNESSetFunction ignores null arguments.

AI assistance: I used Claude Code (Opus 5) to draft parts of this PR. I reviewed, edited, tested, and take responsibility for the final contribution.

[jhale]

This was essentially created by Opus by being told to follow our Python NonlinearProblem and the style/intent of the two other PETSc solver wrappers.

I think NonlinearProblem should be in fem::petsc - this matches the Python interface too.

[jhale]

This is ready for review now - might take some small lessons back to the Python version once this is merged.

[jhale]

It's unclear to me the utility of returning this iteration count, same comment for the KrylovSolver.

[garth-wells]

Fine to change.

[garth-wells]

Fine to change.