Improve batch dimension handling

These are issues left over from the hierarchical modelling project

[ ] Consider leaving the batch dimension in the output from the JaxSolver and process it later with evaluator.create_dataset_from_dict
[ ] Create a batch dimension selector, that constrains the batch (replicate) dimension to the batch IDs specified in sim.coordinates (even though the observations may contain more IDs). This happens during dispatch
[ ] Remove definition of the batch dimension from the solver to the Datastructure of the simulation class if feasible. This has the advantage, that a simple flag can be set and can then also be used for early and informative warnings, but I should be sure that this integrates well with hiearachrical modeling before implementing it