fix(llm_judge): OpenAI base URL, max_completion_tokens for reasoning models, PIL image support

[Luodian]

Summary

• Fix a malformed OPENAI_API_URL default (.../v1/chat/completions/v1) that was being passed straight into the OpenAI(base_url=...) client; now stores a clean base URL (default https://api.openai.com/v1) and appends /chat/completions only for the raw-requests fallback path.
• Apply the same base-URL fix to the async provider, and additionally pass base_url into the AsyncOpenAI(...) client construction -- previously a custom OPENAI_API_URL override was silently ignored by the async client (only the aiohttp fallback respected it).
• Use max_completion_tokens instead of max_tokens for gpt-5*/o1*/o3*/o4* models in both the sync and async providers, via a shared build_sampling_kwargs helper (single source of truth so the two can't drift). For those reasoning models the helper also omits temperature, which they reject at non-default values.
• Accept PIL.Image.Image objects directly in OpenAIProvider._add_images_to_messages (in addition to file-path strings and pre-encoded bytes), reusing the existing encode_image_to_base64 helper which already supports Union[Image.Image, str].

Compatibility note

OPENAI_API_URL now means a base URL (e.g. https://api.openai.com/v1 or your proxy's base). If you previously set it to a full .../chat/completions endpoint, drop that suffix — the SDK client appends the route itself.

Test plan

• python3 -m py_compile on changed files
• uvx ruff check -- clean
• Ephemeral env import + runtime check: default/custom OPENAI_API_URL resolves correctly for both sync and async clients and fallback paths; PIL image branch produces a correct data:image/jpeg;base64,... URL; existing bytes-branch behavior unchanged
• Payload matrix: gpt-4o -> max_tokens + temperature; gpt-5-mini / o3 -> max_completion_tokens only (no temperature, no max_tokens); sync and async providers verified to share the same helper object

Review pass (2026-07-06)

Self-review caught that the original commit applied the max_completion_tokens switch to the sync provider only — the async provider (the default concurrent path) still sent max_tokens, and both providers always sent temperature. Follow-up commit centralizes the logic in build_sampling_kwargs (used by both) and fixes stale images type hints.

Follow-up candidate (out of scope here): azure_openai.py / async_azure_openai.py have the same latent issue (always send temperature + max_tokens); the shared helper is import-ready for them.