refactor: extract EmbeddingService interface (PR-A of #1)

[42U]

Summary

Pure refactor in preparation for issue #1 (configurable embedding providers). Sets up the interface so a second provider can plug in without touching call sites.

• EmbeddingService becomes an interface with providerId and dimensions fields
• Current BGE-M3 implementation renamed to LocalEmbeddingService (providerId: "local-bge-m3")
• New createEmbeddingService(config) factory at the single construction site
• All 27+ files importing EmbeddingService as a type continue to work unchanged

No behavior change. The providerId field is groundwork for the next PR, which adds the embedding_provider column to vector-bearing tables so two providers can never silently mix in the same HNSW index.

Rollout context

This is the first of four PRs landing #1:

• PR-A (this one): provider abstraction
• PR-B: schema, provider tagging, search-time filter
• PR-C: openai-compat provider (covers OpenAI, Azure, Together, Ollama, vLLM, etc. by baseURL)
• PR-D: re-embed migration tool

Test plan

• npm test — 439/439 passing
• Typecheck shows no new errors in changed files (pre-existing peer-dep noise unchanged)
• Manual: existing local BGE-M3 setups continue to work after this lands

Refs #1