fix: PromptOptimizer inherits LLM provider/model from executor's judge when not explicitly set

[amethystani]

Fixes #1856.

When PromptOptimizer runs its internal optimization step (e.g. FeedbackStrategy calling SimplePromptOptimizer to generate an improved prompt), it calls context.llm_wrapper. That wrapper was always built from OptimizerConfig.provider and OptimizerConfig.model, which hardcoded defaults of "openai" / "gpt-4o-mini". So any user whose judge used a non-OpenAI provider (e.g. Vertex AI) hit:

OpenAIError: The api_key client option must be set either by passing api_key to the client
or by setting the OPENAI_API_KEY environment variable

even though they never configured OpenAI anywhere.

Root cause: OptimizerConfig.provider and model defaulted to hardcoded OpenAI strings, so the optimizer always used OpenAI for its own completions regardless of what the user's executor was doing.

Fix:

• OptimizerConfig.provider and model now default to None instead of "openai" / "gpt-4o-mini".
• OptimizerContext gains resolve_provider_model() which checks the executor's judge for provider/model when the config fields are None, then falls back to "openai" / "gpt-4o-mini" for backward compatibility.
• llm_wrapper uses resolve_provider_model() rather than reading config directly.
• BlankLLMJudge._build_judge() raises a clear OptimizationConfigurationError when provider/model are not set (it cannot inherit from an executor since it is the executor creating the sub-judge; users must pass provider= and model= to PromptOptimizer() explicitly).

Before (broken):

optimizer = PromptOptimizer("run", strategy="feedback")
await optimizer.arun(executor=judge_with_vertex_ai, ...)
# → OpenAIError even though no OpenAI configured

After (fixed):

optimizer = PromptOptimizer("run", strategy="feedback")
await optimizer.arun(executor=judge_with_vertex_ai, ...)
# → uses vertex_ai/gemini-2.5-flash for both evaluation and optimization

Users who want a different provider for the optimizer than their judge can still pass provider= and model= explicitly to PromptOptimizer().

Three new unit tests cover the three paths (explicit config, inherited from judge, default fallback).

[amethystani]

@elenasamuylova — tagging you as the maintainer of PromptOptimizer. The fix makes OptimizerConfig.provider/model optional so the optimization step automatically inherits the provider from the executor's judge (e.g. Vertex AI) when not explicitly set, rather than always falling back to OpenAI. Three unit tests cover all three paths. Let me know if the API direction looks right.