Kimi K3 Performance Optimizations in vLLM: The Road to 2.8× Throughput

Kimi K3 serving optimizations across scheduling, KDA prefix caching, ReplaySSM state recovery, PD disaggregation and state offload, parallelism, MoE, and GPU ke