[BUG FIX]: prevent EngineCore crash when Qwen TTS Base task is missing ref_text

[teith]

A single malformed request to /v1/audio/speech with task_type=Base and no ref_text kills the entire EngineCore. All subsequent requests — including valid ones — fail with EngineDeadError.

Reproduce:
serve:
vllm-omni serve Qwen/Qwen3-TTS-12Hz-1.7B-Base --omni --trust-remote-code --host 0.0.0.0 --port 8000 --enforce-eager

from openai import OpenAI
client = OpenAI(api_key="none", base_url="http://localhost:8000/v1")
client.audio.speech.create(
    model="Qwen/Qwen3-TTS-12Hz-1.7B-Base",
    voice="clone",
    input="Hello",
    extra_body={
        "task_type": "Base",
        "ref_audio": "https://cdn-media.huggingface.co/speech_samples/sample1.flac",
    },
)
# Engine is now dead. Every subsequent request returns EngineDeadError.

Root cause: _build_prompt_embeds() raises ValueError when ref_text is missing in ICL mode. vLLM v1 treats any worker exception as fatal.

Fix:
serving_speech.py: validate ref_text presence for Base task before the request reaches the engine (returns HTTP 400)
qwen3_tts_talker.py: fall back to x-vector-only mode instead of raising, matching the existing non-ICL codepath (protects offline inference and other entrypoints)

[yenuo26]

Do we need to add test cases for this scenario? @linyueqian

[linyueqian]

Do we need to add test cases for this scenario? @linyueqian

yes, it would be great.

[linyueqian]

LGTM

[linyueqian]

fix pre-commit please

[linyueqian]

fix dco please

[lishunyang12]

left a minor comment, otherwise looks good

[hsliuustc0106]

fix ci