Allow metrics to return multiple scores

[MatsErdkamp]

Draft PR aims to solve this issue. Allows subscores to be defined in metrics. Declared subscores can be seen by optimizers. An upstream issue that requires this change can be found in the GEPA repo

Example code

    def compute_overall_score(gold, pred, trace, pred_name=None, pred_trace=None):
        metrics = compute_metrics(gold, pred, trace, pred_name, pred_trace)
        quality = subscore("quality", metrics.quality)
        leakage = subscore("leakage", 1.0 - metrics.leakage)

        return (quality + leakage) / 2.0

In the future it would be nice to also adapt MLflow.dspy to autolog the subscores as evaluations.

Would love to have discussions on this draft PR about the implementation. There are some quite big changes that warrant some discussion over syntax, naming and implementation