Bug fixes in ridge tests and reduces dependency on solver

[agoscinski]

first commit is fixing several issues with ridge tests:

• the seed for rng was just set once, that means changing the order of the tests changed the tested matrices in the tests which made debugging hard
• In the tests sometimes np.random.normal was used, but to use the seed one has to use self.rng.normal
• In test test_consistent_weights_scaling the regularization was multiplied with a zero array always, so I removed the parameter and just use zero now
• move documentation of tests

second commit is reducing the dependency on the type of solver by:

• changing tests to
  • Tests without regularization should be high accurate wrt w_exact as long as samples >> rank(X)
  • Tests with regularization should be highly accurate between solvers if num properties is very small otherwise only approximation

second commit also unifies accuracies

• approx: atol=1e-13, rtol=1e-8
• high accuray: atol=1e-13, rtol=1e-8

---

📚 Documentation preview 📚: https://equisolve--30.org.readthedocs.build/en/30/

[PicoCentauri]

Thanks Alex for taking care of these tests! I have just a minor comment but they are already MUCH cleaner!