Feature: Artifact side channel responses for LLM feedback

[jvm123]

Implements the artifact side channel responses feature also for LLM feedback.

• Any LLM feedback response that is not numeric is classified as artifact
• _llm_evaluate returns an EvaluationResult() to evaluate_program()
• evaluate_program() merges the responses

=> LLM feedback strings can contain improvement suggestions that will be available for program improvement

To use this, it is sufficient to use as evaluation prompt something like

Return your evaluation as a JSON object with the following format:
{{
    "metric1": score1,
    "improvement_suggestion": "..",
}}