fix(llm-bench): count inference requests in failure checks

[kiteretsu903]

I found that the failure check counts synthetic metric events as requests, which can make an unreliable endpoint pass. For example, I reproduced a run with 100 inference requests: 90 succeeded and 10 failed. The successful requests also produced 1,710 measurement events. The check divided the 10 failures by all 1,810 events, reporting 0.55% instead of 10% and passing the default 1% limit.

I changed the check to count inference POST requests. That run now correctly fails. I also added explicit success/failure counts, retained summaries for failed runs, prevented duplicate worker output, and preserved CSV percentile columns when the first run has no samples.

---

Note

Low Risk
Benchmark tooling only; changes exit-code semantics and CSV columns without affecting inference serving paths.

Overview
Fixes --max-fail-ratio so it reflects inference failures only, by summing Locust stats for POST requests instead of all recorded events (including synthetic METRIC latency samples). Runs with zero inference POSTs now fail explicitly; failed runs still emit the final summary and CSV row.

The printed summary and --summary-file CSV add Total/Successful/Failed Requests and Failure Ratio. Distributed workers skip the quit handler so only the master evaluates failure and writes output. CSV handling keeps percentile columns (blank when empty), preserves headers when appending (warns and drops unknown new columns), and documents the distinction between POST outcome counts vs existing Num Requests / Qps (completed latency samples).

^{Reviewed by Cursor Bugbot for commit e27d584. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit e27d584. Configure here.}