[Feature] Semantic Validation for the EvalEvalBot

[nelaturuharsha]

I think the easiest option would be for batch processing by a small light-weight LLM (a quantized version is fine that can run on CPU) which is able to process one record at a time - as a prefill task and is able to provide summarized feedback. We can iterate on a system prompt which would allow for an 'agentic' pipeline where the same model can perform different stages of semantic validation (controlled by sub-prompt).

I propose we can also look into https://arxiv.org/abs/2507.19457 to make it better aligned.