Implement LLM check for evaluation

[DachengLi1]

Currently, we follow qwen-math github to parse the evaluation logic. However, many are false negatives - the responses are mostly correct but wrongly parsed.

We should use LLM to check the response.