NVRxStragglerDetection outputs nans as GPU performance scores

[OlegSudakov]

Describe the bug

NVRxStragglerDetection outputs nans as scores. Tested on 2x4 GB200 nodes. Example from logs:

INFO:megatron.bridge.NVRxStragglerDetection:
GPU individual performance:
  Rank=7 Node=<NODE_ID> Score=nan
  Rank=6 Node=<NODE_ID> Score=nan
  Rank=5 Node=<NODE_ID> Score=nan
  Rank=4 Node=<NODE_ID> Score=nan
  Rank=3 Node=<NODE_ID> Score=nan
  Rank=2 Node=<NODE_ID> Score=nan
  Rank=1 Node=<NODE_ID> Score=nan
  Rank=0 Node=<NODE_ID> Score=nan

INFO:megatron.bridge.NVRxStragglerDetection:gpu_relative_perf scores: {'gpu_relative_perf/min': nan, 'gpu_relative_perf/median': nan, 'gpu_relative_perf/max': nan}

Steps/Code to reproduce bug

Attaching .py script and .sh Slurm launch script. Tested on nvcr.io/nvidia/nemo:25.11.nemotron_3_nano container.

Expected behavior

Straggler detection outputs correct GPU performance scores.

slurm_launch_gb200.sh

straggler_detection_pretrain.py

Additional context

Add any other context about the problem here.