refactor(face-embedding): fix BGR/RGB bug and improve embedding pipeline

[merakleee]

What changed

• Fix silent accuracy bug: InsightFace expects BGR, not RGB
• Replace sequential enrollment loop with asyncio.gather (parallel)
• Add detect() method — removes direct model access from service layer
• Pick best face by det_score instead of always taking faces[0]
• Extract _embed_payload() for readability and testability