WKG-80: Allow the user to config the answer correctness prompt

[nelly-hateva]

• Adds configurable answer-correctness prompt via:

  answer_correctness:
    prompt: ...

• Moves the answer-correctness console script implementation from graphrag_eval/answer_correctness.py to graphrag_eval/cli/answer_correctness.py.

• Renames the prompt template variable from {candidate_answer} to {actual_answer}.

• Changes answer correctness parsing to raise ValueError instead of returning an error string.

• Answer correctness evaluator is no longer a singleton

• Changes GenerationConfig defaults:

  • temperature now defaults to 0.0
  • max_tokens is now optional and defaults to None

• Update 3rd party libraries to newest versions due to security vulnerabilities (aiohttp, langchain, pydantic-settings, langsmith)

• Updates docs and tests accordingly.

[pgan002]

Looks good. A couple of minor issues.

[pgan002]

As we discussed in the TTYG meeting, I think we should revert this functionality if we add user configurability of metrics to compute. I think this is the better approach. We should keep the CLI code separation and other improvements.

[nelly-hateva]

As we discussed in the TTYG meeting, I think we should revert this functionality if we add user configurability of metrics to compute.

But they will not be 1:1, because custom metrics can't define how to parse the response and compute something, so if we want to define the answer correctness as custom metric, we will have to prompt the LLM to calculate precision, recall, and f1, while for the built-in metric, we parse the response and calculate these with python.

[pgan002]

OK, please re-request review after you address the comments.