feat(inference): support extra generation model settings

[Teingi]

Which issue or RFC does this PR close?

Closes #1393.

Related to #1379, which covers the broader workload-specific endpoint, header, embedding, and reranking configuration.

Rationale for this change

Some OpenAI-compatible generation endpoints require provider-specific request fields that cannot be expressed in a Pydantic AI model identifier. For example, a compatible Qwen deployment may require chat_template_kwargs.enable_thinking=false.

What changes are included in this PR?

• Add JSON-compatible generation_model_settings to InferenceConfig and Server environment loading.
• Forward the settings to extraction, Experience and Skill generation, Handoff generation, optional LLM reranking, and generation readiness.
• Preserve PowerContext-owned bounds by overriding readiness max_tokens to 1 and reranking temperature to 0.
• Reject settings without a generation model and keep credentials/static headers at the provider boundary.
• Document extra_body.chat_template_kwargs.enable_thinking in the environment example and English/Chinese inference documentation.
• Add focused configuration and request-propagation regression coverage.

Are there any user-facing changes?

Server operators can now set provider-specific generation parameters with one JSON environment variable:

export POWERCONTEXT_SERVER_INFERENCE_GENERATION_MODEL_SETTINGS='{"extra_body":{"chat_template_kwargs":{"enable_thinking":false}}}'

The default remains an empty object, so existing deployments keep their current behavior. Only settings supported by the selected Pydantic AI model/provider have an effect. This setting does not replace provider credential configuration.

How was this change tested?

• make check
• make test (953 passed, 9 skipped)
• make docs-test
• Focused tests for nested environment parsing, validation, generator settings, and readiness propagation

AI usage statement

OpenAI Codex with GPT-5 was used for implementation, tests, documentation, and review.

[PsiACE]

Could we either document the precedence between PowerContext-owned settings and extra_body, or enforce an explicit conflict policy? merge_model_settings only handles top-level keys.

[PsiACE]

For example, max_tokens=1 can conflict with extra_body.max_output_tokens=100, leaving the effective value dependent on the provider/SDK merge order.

[PsiACE]

totally lgtm