Get rid of model access in rasa.core.test

[wochinge]

Description of Problem:
The rasa.core.test module has several paths where it directly access attributes on the Agent or even lower. This has the disadvantage that the test code is really tied to the specific model architecture. With the revamp we want to reduce these ties. This will help us to keep the interface of Graphprocessor small and explicit and will reduce the amount of breaking things when we start to implement the actual graph.

Related implementation proposal page

Currently the following pieces in the code directly access Agent stuff which we don't want it to access

1. Tokenizing messages for the entity evaluation: We can skip this case for now ⚠️ We handle it once we have handled the prediction metadata task which adds the token to the output of every parsed user message.
2. Determining whether we should predict another action: My idea was to change this check to raise an exception ActionLimitReached when we call processor.predict_next_action instead of having a separate method which requires callers to know the logic
3. 2nd part of determining whether predict another action: Same we can handle this with an exception as well in the test module

Overview of the Solution:
I outlined some proposed ways of getting rid of these accesses. Anything else which doesn't require an additional method on the GraphProcessor is also fine.

Definition of Done:

• The described locations 2) and 3) no longer replicate logic from the processor