feat(embed): retry transient embedding errors

[RomneyDa]

Summary

• Classify OpenAI embedding failures via a typed APIError (status / type / code / Retry-After).
• Retry transient ones (429, 5xx, network timeouts) with Retry-After-aware exponential backoff + jitter; longer base for overloaded_error. insufficient_quota, 4xx, and ctx errors surface immediately.
• Replace abort-on-first-error with a per-batch retry queue: each batch retries once with fresh backoff and the rest keep going. Final run status is success / partial / error / cancelled, with retries and per-batch failure metadata persisted to stats_json.

Test plan

• go test ./...
• Covers Retry-After (seconds + HTTP-date), insufficient_quota no-retry, overloaded_error backoff, ctx cancellation, partial-success runs, cancelled runs.