Feat/multi node rank collective report

[mvstrauss]

Summary

• Add --trace_glob + --rank_regex support to generate_multi_rank_collective_report_pytorch.py to handle tensorboard-style / multinode trace filenames and nested directories.
• Add optional node-aware reporting via --gpus_per_node, producing node-span (intra_node vs inter_node) summary sheets.
• Add minimal progress/status prints (with elapsed time) during trace resolution, loading, and report generation to avoid long “silent” runs.

Details

• New input mode: --trace_glob (recursive glob) with rank extraction using --rank_regex.
• Node-span summary sheets (when --gpus_per_node is set):
  • nccl_summary_long_node_span
  • nccl_summary_implicit_node_span
• Default output path is derived from the resolved trace file paths (works for glob/pattern modes).
• Documentation updated in docs/generate_multi_rank_collective_report_pytorch.md.

Test plan

Usage example for summarizing profiles from a 2 node experiment, 8 GPUs per node.

python -m TraceLens.Reporting.generate_multi_rank_collective_report_pytorch \
  --trace_glob "/path/to/your/run/tensorboard/**/**.pt.trace.json.gz" \
  --rank_regex "rank\\[(?P<rank>\\d+)\\]" \
  --world_size 16 \
  --gpus_per_node 8 \
  --detailed_analysis \
  --output_xlsx_path "/path/to/output/nccl_analysis_report.xlsx"