System-aware P/D decision making to adapt to workload shifts

[RishabhSaini]

What would you like to be added:
Enhance P/D decision logic to consider system state (queue depths, worker load, predicted latencies, etc.), not just non-cached token counts. Allow the scheduler to incorporate dynamic metrics when deciding whether to use disaggregated prefill/decode.

Why is this needed:
Current token-threshold-based decisions can't adapt to provisioning or workload changes. Example: if workload shifts from 10k→20k prompt tokens and prefill workers become overloaded, the scheduler should automatically reduce P/D usage to avoid bottlenecks. System-aware decisions would enable dynamic adaptation without manual threshold tuning.