bench: report cross-rank timing stats

[mawad-amd]

Summary

• Report timing from all ranks instead of only rank 0. Each rank computes the median of its iteration times, then dist.all_gather collects every rank's median to rank 0.
• Headline GPU Time is now max across ranks (the true collective latency — slowest rank determines when the operation is done).
• Two new counter columns: min_ms (fastest rank) and skew_% ((max - min) / max * 100, straggler penalty as a percentage).
• Supports both NCCL (CUDA tensors) and gloo (CPU tensors) backends for the gather.

Example output on MI308X ×8:

num_ranks     M     N     dtype   variant  GPU Time (ms)  BW (GB/s)  min_ms  skew_%
        8  1024  1024  bfloat16  two_shot          0.217       16.9   0.215   0.953
        8  4096  1024  bfloat16  two_shot          0.272       53.9   0.270   0.758

Test plan

• Ran bench_all_reduce.py on MI308X ×8 with --axis_num_ranks=8 --axis_M=1024,4096 --axis_N=1024 --axis_dtype=bf16
• Verify JSON and CSV output formats include new counters
• Test with --axis_num_ranks=2,4,8 to confirm multi-launch still works

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Updates the benchmark runner to report cross-rank timing statistics so the headline “GPU Time” reflects true collective latency (slowest rank), and exposes cross-rank variability via new counters.

Changes:

• Compute per-rank median iteration time and aggregate timing stats across ranks.
• Define “GPU Time” as max rank median, and compute bandwidth/TFLOPs from that value.
• Add min_ms and skew_% counters to quantify straggler impact.