[TypeSpec Authoring][Benchmark] Run evals with 2 trials

[haolingdong-msft]

Summary

Adds multi-trial support to the azure-typespec-author skill evaluation suite using Vally's native config.runs mechanism, and updates the result reporter to aggregate trials with pass@k semantics and surface each trial's per-case status.

Motivation

Single-run evals are flaky: a transient miss on one stimulus drops the suite pass rate and produces noisy red builds. Running each stimulus twice and judging pass@k separates real capability failures (fail every trial) from flakiness (passes at least once).

Changes

• config.runs: 1 → 2 across all 28 evals/*.eval.yaml and 3 suites/{forced,trigger,no-skill}.eval.yaml specs — every stimulus now runs twice per evaluation.
• scripts/report-eval-results.sh updated to:
  • Group the now-multiple results.jsonl records per stimulus by stimulusName (Vally emits one record per trial when runs > 1).
  • Apply pass@k: a stimulus passes if any trial passes; only stimuli that fail every trial raise an ADO ##vso[task.logissue type=error].
  • Print a per-trial headline, flag [FLAKY] stimuli (pass some-but-not-all trials), and list each case's status across trials.

Sample output

--- Suite versioning trial results ---
trial 1 (10/11)
trial 2 (9/11)
fail in both trials: 1

Pass@k: 10/11 (90%), 1 flaky
  010 [PASS] trial1:PASS trial2:FAIL tool-calls:pass
  011 [FAIL] trial1:FAIL trial2:FAIL

Validation

• All 32 eval/suite specs confirmed at runs: 2 (0 leftover runs: 1).
• bash -n clean on the report script.
• Report logic verified against synthetic 2-trial data: flaky → PASS (flagged), fail-both → FAILED + error, correct trial counts and pass@k rate.

⚠️ Follow-up (not in this PR)

2 trials roughly doubles runtime. The per-suite TimeoutMinutes in eng/pipelines/azure-typespec-author-benchmark.yml (versioning 50, armtemplate 40, lro 25, decorators 15, warning 25) may need bumping.