Add new metric functions: LLMGEvalScore and ChatLLMGEvalScore

[m-ast]

Summary

• Add two metric functions: LLMGEvalScore and ChatLLMGEvalScore
• These two metics calculate a weighted average score for evaluation of lm_output.

Implementation

• Create flexeval/core/metric/llm_geval_score.py
  • Implement LLMGEvalScore and ChatLLMGEvalScore which inherit Metric
  • Note that these metrics can be used only for HuggingFaceLM and VLLM, implemented batch_compute_log_probs. I plan to implement it for OpenAIChatAPI and OpenAIChatBatchAPI afterwards.
• Create tests for the script.