Add basic building blocks for a PETSc solver

[max-models]

This is a redo of some of the PETSc testing I did long ago.

Since psydac has a mat_topetsc() and vec_topetsc(), using petsc for at least some solves should be pretty straightforward.

Todo:

• Create a PETScSolver class
• Create an abstract solver class
• Support mass matrix solves
• Full support, including composed operators
• Integration in the struphy models
• Integration via struphy simulation API

---

Notes

• No speedup with petsc for the mass-matrix case, feectools' compiled matvec is cheap enough that even tens of thousands of extra CG iterations stay cheaper than PETSc's one-time GAMG setup + per-solve Python-level vector conversion. Maybe if the same matrix is solved repeatedly with different right-hand sides, PETSc only pays the GAMG setup cost once.