[Router] Log every request at one site; derive its outcome from the final status (3/3) (#39465)

Co-authored-by: Kangyan Zhou <kangyan.zhou@radixark.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>

experimental/sgl-router/monitoring/grafana-dashboard.json

@@ -100,7 +100,7 @@
     {
       "type": "stat",
       "title": "Error ratio",
-      "description": "Share of dispatches whose outcome=error.",
+      "description": "Share of dispatches that failed on the WORKER's side (outcome=error). Excludes client_error (4xx), backpressure (429/503) and cancelled, so a saturated fleet or a bad client does not read as a worker fault.",
       "datasource": {
         "type": "prometheus",
         "uid": "${datasource}"