feat(gemini): migrate to google-genai SDK, keep gemini_api alias

[Luodian]

Gemini: migrate to the google-genai SDK

Public shipped the deprecated google-generativeai backend (gemini_api.py). This PR modernizes the Gemini integration to the current google-genai SDK.

What changed

• New backend lmms_eval/models/simple/gemini.py (Gemini, registered as gemini): uses google.genai, with thread-safe video/audio file upload + cache, adaptive concurrency (via model_utils.concurrency_control), prefix-aware request queue, safety-block salvage, and token/usage accounting (model_utils.usage_metrics).
• New chat backend lmms_eval/models/chat/gemini.py (Gemini(GeminiSimple), is_simple=False): converts the ChatMessages protocol to Gemini-native contents and extracts system messages into system_instruction.
• Registry (lmms_eval/models/__init__.py): gemini registered as both a simple and a chat model; gemini_api kept as an alias of gemini (MODEL_ALIASES["gemini"] = ("gemini_api",)) so existing --model gemini_api commands keep working.
• Removed the deprecated lmms_eval/models/simple/gemini_api.py (the CLI name survives via the alias).
• pyproject.toml: gemini extra google-generativeai -> google-genai>=1.0.0.
• New lmms_eval/models/model_utils/parallel.py: small generic ParallelPlan / resolve_local_rank_device launcher-parallelism helper (stdlib + lazy torch).

Compatibility

• --model gemini and legacy --model gemini_api both resolve to the new Gemini class (verified via the registry manifest + alias resolution).
• Backend switch: because gemini is now registered as a chat model too, --model gemini / --model gemini_api default to the chat backend (same convention as openai); pass --force_simple for the simple path. The interleave model_arg applies to the simple path only.
• Default model: model_version defaults to gemini-3.5-flash (the old backend defaulted to the now-deprecated gemini-1.5-pro); pin model_version=... in --model_args if you need a specific model.
• Concurrency default: num_concurrent=32 with linear backoff (the old backend was serial). On low-quota keys, lower it or set adaptive_concurrency=True.

Notes

• pip install -e ".[gemini]" now pulls google-genai>=1.0.0 instead of the deprecated google-generativeai.
• parallel.py has no in-tree consumer yet; it is included as a reusable helper (feat(models): add diffusers-based world-model generation backends (Wan 2.1/2.2, Cosmos, LTX-Video, MAGI-1, fal.ai) #1389 inlines a minimal copy — dedupe at merge, whichever lands second).
• Known follow-up: file uploads run on the dispatch thread (the thread pool parallelizes generate_content only), so long video uploads serialize; upload-in-worker is a natural next step.

Review pass (2026-07-06)

Two self-review commits:

• Upload polling compared str(FileState.ACTIVE) ("FILESTATE.ACTIVE") against "ACTIVE" — it never matched, so every unique video/audio upload burned the full poll budget (~120s) and FAILED uploads were never surfaced. The loop now normalizes the enum via its name and raises on FAILED.
• Audio file paths were always tagged mime_type="audio/wav". Both backends now map extensions to Gemini-canonical MIME types (.mp3 -> audio/mp3, .flac -> audio/flac, ..., fallback mimetypes.guess_type then audio/wav).