[CI][Perf] Add high-load stress phase for Qwen3-TTS daily perf

[linyueqian]

Summary

Daily DFX TTS perf currently caps at max_concurrency=8 in the throughput regime, so high-load TTFA tail regressions (e.g. the cross-request Code2Wav batching gap that #3163 proposes to fix and that #3221's Triton stack already demonstrates a fix for) are invisible to nightly CI.

This PR adds a stress phase to test_qwen3_tts_customvoice for both default_voice and voice_design, mirroring the open-loop pattern already used by test_qwen_omni.json:

• num_prompts: [100], request_rate: [2.0] (open-loop, ~2 req/s offered for ~50 s of wall time)
• Same prompt source as the existing throughput phase, so no new dataset deps
• Baselines intentionally loose (median TTFA 3.0–3.5 s, median RTF 0.25–0.30, audio_throughput floor 4.0 audio-s/wall-s) so it alarms only on real regressions and can be tightened once we have a few nightly runs

eval_phase: \"stress\" is metadata only — run_benchmark.py already lists eval_phase in exclude_keys, so no script change is needed.

Why this matters

On H20 (single H20-3e), the gap between current main and a stack with cross-request codec batching shows up exactly in this load region:

Concurrency	main req/s	main TTFA p95	PR #3221 req/s	PR #3221 TTFA p95
8	2.29	386 ms	4.41	397 ms
16	2.44	4437 ms	4.52	258 ms
32	2.47	8626 ms	6.73	463 ms

main saturates at ~2.47 req/s starting at c=8, so any offered load above that queues hard. The new entry sits at offered rate=2.0 (~80 % of main's sustainable rate, ~30 % of #3221's) — that's the regime where regressions in scheduler / codec batching are loudest.

Test plan

• python3 -c \"import json; json.load(open('tests/dfx/perf/tests/test_tts.json'))\" parses cleanly
• tests/dfx/perf/tests/test_runner_metadata.py (already excludes eval_phase) still passes
• First nightly run produces baseline metrics for the new stress entries; tighten the floors in a follow-up PR
• After [RFC]: Cross-request batching for Qwen3-TTS Code2Wav stage to fix TTFB scaling under concurrency #3163 lands, re-check the median TTFA / RTF on the stress entries to confirm the win is captured by daily

cc @ischencheng (re #3163), @vklimkov-nvidia (re #3221), @hsliuustc0106

[hsliuustc0106]

Review summary

Validated:

• All gates pass (DCO, pre-commit, build 3.11/3.12, docs)
• eval_phase is in exclude_keys at run_benchmark.py:336 — no script change needed
• New entries mirror the existing throughput/latency structure exactly (same percentile-metrics, dataset_path, backend, consistent per-task config)
• Baselines are intentionally loose, documented as such, with a plan to tighten after nightly data
• Benchmark evidence in the PR body is thorough: concrete concurrency/throughput/TTFA table showing exactly the gap this stress phase will catch

No blocking issues. The change is well-scoped (config-only, +36 lines), well-justified, and well-evidenced. LGTM.