Triton TensorRT-LLM backend README - the six nv_trt_llm_* label-switched metric families

README.md 'Triton Metrics' section on main, 2026-10-04, with sample exposition output. The backend contributes six metric families, all prefixed nv_trt_llm_, each carrying every number as a LABEL rather than as a metric name: nv_trt_llm_request_metrics{request_type="waiting"|"context"|"scheduled"|"max"|"active"}, nv_trt_llm_runtime_memory_metrics{memory_type="pinned"|"gpu"|"cpu"}, nv_trt_llm_kv_cache_block_metrics{kv_cache_block_type="fraction"|"tokens_per"|"used"|"free"|"max"}, plus inflight_batcher_metrics, general_metrics and disaggregated_serving_metrics. The sample output confirms model="tensorrt_llm", version="1" are also labels. This is why the LLM metrics a Triton+TRT-LLM deployment has to be queried by filtering a label rather than by naming a metric, and it is a structurally different shape from every engine-native metric set in this repo.