feat(inference): support workload-specific model endpoints

[PsiACE]

Which issue or RFC does this PR close?

Closes #1379.

Rationale for this change

Generation, embeddings, and LLM reranking may use different model services and require different routing or authentication headers.

What changes are included in this PR?

• Add workload-specific models, base URLs, static headers, and model settings.
• Configure sensitive headers on provider clients instead of request settings.
• Keep LLM reranking on the regular model interface.
• Document provider-owned operation paths and configuration limits.
• Add acceptance coverage for separate generation, embedding, and rerank services.

Are there any user-facing changes?

New inference environment variables are available. Existing configurations keep their current behavior.

Custom endpoints and headers require an explicit OpenAI- or Anthropic-compatible model identifier. Arbitrary operation-path rewriting is not supported.

How was this change tested?

• make check
• make test
• make docs-test
• Local acceptance services for generation, embeddings, and LLM reranking

AI usage statement

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