Less error-prone design of linear solver state

[JonathanWenger]

Is your feature request related to a problem? Please describe.
Currently a lot of care has to be taken when caching quantities during an iteration of the ProbabilisticLinearSolver. With the described changes below this can be made less error-prone in the component implementations making use of the state. Ideally, someone who designs a new component should not need to be aware of the details of the cache. The points below can improve this.

Describe the solution you'd like.
The following if implemented would improve the current state implementation:

• Do not make the random number generator / seed part of the state. Reorthogonalization for a stable implementation of the probabilistic linear solver #580 (comment)
• Turnstate.cache into a collections.defaultdict(list) to avoid having to check for initialization of caches Reorthogonalization for a stable implementation of the probabilistic linear solver #580 (comment)_
• Introduce getters and setters for cached quantities to ensure they are only computed once Reorthogonalization for a stable implementation of the probabilistic linear solver #580 (comment)