Metrics and Algorithms to Observe and Classify LLM Backend Load

[rootfs]

Is your feature request related to a problem? Please describe.
Explore and expose additional Prometheus metrics from the Semantic Router to describe workload characteristics (prompt/response sizes, category distribution, cache hit ratio) and backend load (endpoint utilization, token throughput, TTFT, TPOT). These metrics enable the control plane to implement algorithms that adapt router configs dynamically for latency, accuracy, and cost objectives.

[Xunzhuo]

/area observability core
/priority P1
/milestone v0.1

[tao12345666333]

Please assign this issue to me.