Pluggable embedding providers (VoyageAI, OpenAI, Ollama)

[GoodbyePlanet]

server/embeddings/base.py:7 already defines an EmbeddingProvider Protocol, but only
JinaEmbeddingProvider is wired up (server/embeddings/jina.py). claude-context supports OpenAI /
VoyageAI / Gemini / Ollama and lets users pick. Notably, Voyage voyage-code-3 outperforms
jinaai/jina-embeddings-v2-base-code on most code retrieval benchmarks, so this is also a quality
lever, not just a flexibility one.

Scope

• Add VoyageEmbeddingProvider (POST https://api.voyageai.com/v1/embeddings, model default
  voyage-code-3, input_type="document" for indexing and "query" for search). Honour batching
  (Voyage caps batch size at 128).
• Add OpenAIEmbeddingProvider (default text-embedding-3-large).
• Add OllamaEmbeddingProvider (default nomic-embed-text, host from env).
• Provider selection in server/config.py via EMBEDDINGS_PROVIDER=jina|voyage|openai|ollama
  (default jina, preserving current behaviour).
• Each provider exposes its native dimensions; EMBEDDINGS_DIMENSIONS becomes optional /
  provider-derived. Make sure the dimension mismatch with an existing Qdrant collection produces a
  clear error, not a silent failure.
• Update README.md env table.
• Tests: mock-based unit tests for each provider's request shape and batching behaviour (respx is
  already in dev deps).

Acceptance

• Setting EMBEDDINGS_PROVIDER=voyage + VOYAGE_API_KEY lets the server index and search without
  running the local TEI container.
• Default behaviour (Jina via TEI) is unchanged when env is unset.
• docker-compose.yaml keeps Jina TEI as the default but documents how to disable it for hosted
  providers.