Use the `Evaluator` class only to construct the function call or even to validate it only

[flo-schu]

The reason is that with passing through of important arguments the overhead of instantiating it again and again is not necessary. The challenge is how to dynamically call it with different arguments.

• [ ] Infer evaluator return dimensions by executing the evaluator once on setup if the return dimensionality is not configured and issue warnings if there is danger of confounding dims

🛑 This should be carefully thought through.

• [ ] write pseudo code to get started
• [ ] Run tests to make sure everything still works

[flo-schu]

This issue came up again, when implementing a more versatile solver. The solver is slow, because it repretitively executes the same code again and again.

This can be drastically improved by outsourcing the repetitive tasks to the evaluator, which is, on top, of that only instantiated once.

The Evaluator

• combines everything that is needed to evaluate a model.
• It is instantiated just before the evaluation of a model is started
• It should be instantiated on every evaluation, because of parallelization, each evaluator instance must be a self contained instance that also carries the results
• However, much of the evaluator will remain fixed for each evaluation:
  • model,
  • post-processing
  • solver,
  • data,
  • indices,
  • data-variables
  • coordinates
  • ode-states
• The arguments that remain free are in fact only model_parameters with the keys y0, x_in and parameters
• For this input, transformations & interactions with parameters should be allowed, or e.g. an x_in could also be sampled from a distribution, ...
• A evaluator "factory" should be constructed before an inference algorithm or similar is started. This should set up the evaluator with all required static information