Wire production AsyncLLM requests into Prometheus /metrics

Problem

The shipped vllm-server --enable-metrics exposes the Prometheus catalog, but production requests run through AsyncLLM while the logger is attached only to the unused synchronous LLMEngine. Live scrapes therefore leave num_requests_running, token counters, request-success counters, and timing histograms at zero even while requests complete.

Reproduction

Run four concurrent /v1/chat/completions requests against a server started with --enable-metrics and scrape /metrics at 100 ms cadence. Requests complete correctly, but all nine aligned samples report num_requests_running=0, num_requests_waiting=0, with token and success counters unchanged.

Scope

Complete W6 in .agents/specs/prometheus-metrics.md:

• attach the existing PrometheusStatLogger to production AsyncLLM;
• build and record IterationStats in the output-handler path;
• make concurrent Record() / Expose() safe;
• preserve the null-logger no-stats path;
• add RED-first async metric and concurrent-scrape tests;
• gate affected async/sync/Prometheus/server targets.

No model kernels, sampling, scheduler policy, or Hermes client changes.

Read .agents/specs/prometheus-metrics.md and trace the production AsyncLLM output-handler path alongside PrometheusStatLogger and IterationStats. Add the async metric recording and concurrency coverage described there, preserving the null-logger path, then run the affected async, sync, Prometheus, and server targets to verify RED metrics and concurrent scrapes.