SGLang cached prompt-token counter

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. sglang:cached_tokens_total additionally carries a cache_source label in {device, host, storage} (or the literal 'total' on the backward-compatibility path), so prompt-cache reuse is decomposable by tier - finer than any other engine in this fleet.