Decoupling linear solvers from PDE solvers

So, this is most likely ready - please review 😉 @oberhuber @hanousek

Impact on compilation times:

For tnl-mhfem: before: 3m49.772s, after: 2m58.645s. But before, I had just one linear solver enabled, now I have all of them 😆

But memory usage during compilation and compilations on CI are not affected much (9m09s vs 7m40s for cuda_Release), probably because there are many things which are not affected by this (e.g. tests).