Cost Observability: Token-Level Cost + FinOps Tagging + Idle GPU Detector

Production LLM TCO'sunu kontrol altına almak: per-request token cost tracking, customer-level FinOps tagging (kimin user'ı kaç token), idle GPU detector (vLLM serving'de utilization %50'nin altına düşerse alarm), cost-per-query trend, alarm thresholds.