Unify evaluation API of model-based and statistical metrics

[tholor]

We recently introduced two different ways to evaluate a pipeline:

1. Statistical metrics
2. Model-based metrics

Let's unify the interface here so that this is a consistent UX.
We already discussed that we will go for the new "model-based interface":

• A single StatisticalEvaluator component
  • One metric per-instance
  • Inputs based on the metric

[shadeMe]

The DeepEvalEvaluator implementation can serve as a blueprint for the statistical evaluator - deepset-ai/haystack-core-integrations#346