Fix: evaluation leaks api_key in metadata; fix llm_config argument in run infer

[xingyaoww]

What is the problem that this fixes or functionality that this introduces? Does it fix any open issues?

• after the global config refactor Remove global args #2760, llm_config inside EvalMetadata no longer hides sensitive info when dumped into JSON.
• the LLM's input keyword argument change from llm_config to config

---

Give a summary of what the PR does, explaining any non-trivial design decisions

• Hide sensitive info when try to dump EvalMetadata into a JSON via model_dump_json
• Fix all occurrence of llm_config= to config= in all run_infer.py

---

Other references

[neubig]

Thanks for cleaning this up!