[Fix] Strip use_vllm/use_verifier eval flags from judge kwargs (OpenAI 400)

[cdpark92]

Bug

run.py forwards the --use-vllm and --use-verifier CLI flags into
judge_kwargs:

# run.py, get_judge_kwargs
if args.use_verifier:
    judge_kwargs['use_verifier'] = True
if args.use_vllm:
    judge_kwargs['use_vllm'] = True

build_judge(**judge_kwargs) only pops model and nproc; everything else
reaches the API wrapper, and OpenAIWrapper.generate_inner places unknown
kwargs directly into the request body:

payload = dict(model=self.model, messages=input_msgs, n=1,
               temperature=temperature, **kwargs)

The real OpenAI API rejects unknown body fields:

400 - Unrecognized request argument supplied: use_vllm

Since even the working() health probe sends this payload, every judge-scored
evaluation fails up front (MMVet crashes on its assert model.working();
MathVista reports 100% judge failure; MCQ answer extraction degrades) whenever
a model is evaluated with --use-vllm and an OpenAI judge.

vLLM-served local judges ignore unknown request fields, which is why the bug
is easy to miss when developing against a self-hosted judge.

Fix

Pop the two eval-only flags in build_judge() before constructing the API
wrapper (+3 lines). This covers every build_judge(**judge_kwargs) call site
at once. The flags remain visible to the dataset evaluate() dispatch, which
reads them from judge_kwargs before build_judge is called — ** expansion
passes a copy, so callers are unaffected.

Reproduction

# any judge-scored benchmark + --use-vllm + a real OpenAI key
python run.py --data MMVet --model <any-vllm-capable-model> --use-vllm
# before: AssertionError: MMVet evaluation requires a working OPENAI API
#         (judge probe rejected with 400 'Unrecognized request argument supplied: use_vllm')
# after: judge calls succeed