Injecting Adrenaline into LLM Serving: Boosting Resource Utilization and Throughput via Attention Disaggregation