Refactor `LinearOperator`s to use `probnum.backend`

[marvinpfoertner]

Is your feature request related to a problem? Please describe.
After merging #581, Normal RandomVariables with LinearOperator covariances will only work in the NumPy backend. Moreover, there is still a lot of LinearOperator-related code in the Normal implementation.

Give an example use case.

• differentiable linear solvers and ODE filters

Describe the solution you'd like.

• refactor LinearOperators into a generic implementation using the backend
• this should be done one operator at a time, starting with LinearOperator and Matrix

Additional context
None