Handle vLLM 0.17 rank trace filenames

[powderluv]

Summary

• match both legacy -rank-0 and current vLLM rank0 torch profiler filenames
• exclude async_llm coordinator traces from per-rank gap analysis discovery
• keep the fallback path limited to non-async trace files

Problem

With vllm 0.17.0, torch profiler outputs rank traces like:

• dp0_pp0_tp0_dcp0_ep0_rank0....pt.trace.json.gz
• dp0_pp0_tp1_dcp0_ep1_rank1....pt.trace.json.gz

and also emits a coordinator trace like:

• ...async_llm....pt.trace.json.gz

GapAnalyzer.detect_trace_files() only matched the older -rank-0 style. That meant rank detection failed, the code fell back to *.json.gz, and async_llm was incorrectly counted as an extra rank.

Validation

Validated against a real vllm 0.17.0 Kimi K2.5 trace directory from an MI355 run:

• before: fallback path picked up 5 traces including async_llm
• after: detector returns exactly 4 rank traces (rank0..rank3)

Notes

I did not add a unit test in this patch because I was validating against the live trace artifacts first. The next clean step would be a small test covering both filename styles and the async_llm exclusion.