Add Ollama-backed LLM and embedding providers

[admin-derinworks]

Support running the RAG service against a local model served by Ollama, so retrieval becomes semantic and POST /answer returns real generated answers instead of the mock echo — without giving up the dependency-free default build.

The EmbeddingProvider and LlmProvider traits (#5) were designed for exactly this: an Ollama backend is a drop-in implementation of each, with no changes to the retriever, generator, or handlers.

Requirements

• OllamaLlmProvider implementing LlmProvider against POST /api/generate ("stream": false), returning the response field
• OllamaEmbeddingProvider implementing EmbeddingProvider against POST /api/embed, batching inputs in one call ("input": [...] → "embeddings": [[...]])
• Gate both behind an optional ollama cargo feature that pulls in an HTTP client (e.g. reqwest), keeping the default build dependency-light
• CLI wiring: serve --llm ollama --llm-model <name> --embed-model <name> [--ollama-url <url>] replacing the hardcoded mocks when the feature is enabled
• Surface backend failures as LlmError::Provider / EmbeddingError::Provider so the existing HTTP error mapping (502 for backend failures) applies unchanged
• Document setup in the README (ollama pull, choosing an embedding model, min_score tuning now that scores are real cosine similarities)

Notes

• One embedding model per index: the vector store pins dimensionality on first insert (e.g. nomic-embed-text is 768-dim), so index and query must embed with the same model.
• Integration tests should run against a live Ollama only when one is reachable and skip otherwise; the mock-based tests remain the CI gate.