Add OpenAI-compatible embedding API support for evaluation

[cindehaa]

Summary

• Tooling only
• Adds a way to evaluate API-based embedding models via generic encoder for OpenAI-compatible providers

Validation

• Ran the full pipeline (embed → retrieve → evaluate → write runfile + results JSON) end-to-end against the live voyage-4-large API on a small langchain slice
• OpenAI client works against Voyage's endpoint, the response is data[].embedding-shaped, and the run produces a valid TREC runfile and metrics JSON