docs: add --tail=-1 to the label-selector log command

Review catch, and it is a defect this PR introduced. kubectl's own help states
the default plainly: 'Defaults to -1 with no selector, showing all log lines
otherwise 10, if a selector is provided.'

So switching from a named pod to a label selector silently capped output at the
last 10 lines per worker, which would omit the 'NCCL INFO ... Using network'
banner NCCL prints during init — the exact line the command greps for. The
check would have reported nothing on a correctly attached job.

Refs #2306

Signed-off-by: Yuan Chen <yuanchen97@gmail.com>

Author: yuanchen8911

docs/user/fabric-attached-training.md

@@ -295,7 +295,9 @@ socket fallback:
 
 ```shell
 # Select workers by label rather than guessing the generated pod name.
-kubectl logs -n <namespace> -c node \
+# --tail=-1 is required: with a selector, kubectl defaults to the last 10 lines
+# and would omit the transport banner, which NCCL prints during init.
+kubectl logs -n <namespace> -c node --tail=-1 \
   -l jobset.sigs.k8s.io/jobset-name=<trainjob-name>,jobset.sigs.k8s.io/replicatedjob-name=node \
   | grep -i 'NCCL INFO.*Using network'
 ```