Refactor the design of the probabilistic linear solver functions and classes

[JonathanWenger]

• Cleanly separate _preprocess_linear_system from _init_solver and make sure only one takes the assumptions on A, i.e. assume_A as an argument.

• Create a common (abstract) class MatrixBasedSolver(ProbabilisticLinearSolver) with generic functions:

  • policy
  • observe
  • step_size / line_search
  • infer_posteriors
  • optim_hyperparams
  • solve
  • ...
    This way we can avoid some code duplication between the different versions of the matrix-based solvers.

[nathanaelbosch]

In the initial pylint PR #150 I found many pylint messages that mostly relate to linalg, and especially matrixbased. I'll post a full list in here since it seems to relate to this PR. If you think that those are better tackled through individual tasks I could also open individual issues:

• too-many-arguments
• abstract-method
• too-many-branches
• arguments-differ
• redefined-builtin
• too-many-lines
• abstract-class-instantiated
• too-complex
• too-many-instance-attributes
• too-many-statements
• attribute-defined-outside-init

These are not all the errors that pylint raises, neither are those errors that only appear in linalg, but I think that most of these should be considered when working on this refactoring, and ideally I would hope for pylint src/probnum/linalg/ to pass with these messages enabled (e.g. through the --enable=<message> argument).

[nathanaelbosch]

Relating to my previous comment: I just opened #158 to track the work on the pylint exceptions for the linalg module.