SGLang per-stage request latency histogram (stage label)

SGLang main, python/sglang/srt/observability/metrics_collector.py, read 2026-10-04. NOT ON BY DEFAULT: nothing is exported unless --enable-metrics (ServerArgs default False), which is also the only thing that mounts /metrics on the server port. Label keys are model_name, engine_type, tp_rank, pp_rank, moe_ep_rank (plus dp_rank under data parallelism, priority under priority scheduling, and anything in extra_metric_labels). Bucket ladders are overridable per histogram via get_histogram_conf_from_env. exponential_buckets(start=0.001, width=1.62, length=30), covering roughly 1 ms to 1191 s. stage names come from the SAME RequestStage enum that names the OTel trace slices, so metrics and traces share one vocabulary - the property no other engine here has.