fix: guard choices[0] and message=None before content access (41 sites, 32 files)

[qizwiz]

Summary

OpenAI-compatible APIs can return empty choices on error or rate-limiting, causing IndexError at choices[0]. Gemini additionally returns choices[0].message = None on PROHIBITED_CONTENT safety filtering (HTTP 200 — no exception raised), causing AttributeError at .content.

This PR adds the comprehensive guard:

if not response.choices or response.choices[0].message is None:
    raise ValueError("LLM returned empty or filtered response")

before every bare choices[0].message.content access across 32 files, 41 sites.

Impact

Affected subsystems:

• lmms_eval/llm_judge/providers/ — OpenAI and Azure OpenAI judge providers
• lmms_eval/models/ — core model implementations (chat/openai, simple/openai, srt_api, gpt4v)
• lmms_eval/tasks/ — 25+ task evaluation utilities across aime, babyvision, gpqa, live_bench, mathverse, mix_evals, mme_sci, openai_math, phyx, tomato, videoevalpro, videomme, and unig2u tasks
• test/models/test_litellm.py

Verification

The Gemini choices[0].message = None crash is a documented production behavior on PROHIBITED_CONTENT safety filtering. Static analysis via pact (Z3 Fixedpoint datalog, llm_response_unguarded rule) found all 41 sites; post-fix scan returns 0 violations.

[kcz358]

Is this file got wrongly modified? The git diff looks not as expected as well as the other utils in mme_sci