[Feature]: Add word-level timestamps to TTS streaming output via shared forced aligner

🚀 The feature, motivation and pitch

Voice-agent use cases (barge-in, mid-utterance interrupt) need word-level timestamps aligned to the streamed audio. Currently the vllm-omni TTS streaming path only returns audio chunks, so an orchestrator can't tell which word the user interrupted on. This applies to every TTS model in the repo: Qwen3-TTS, VoxCPM2, CosyVoice3, FishSpeech, OmniVoice, MOSS-TTS-Nano, Voxtral-TTS.

Proposal: emit (word, start_ms, end_ms) alongside each audio chunk in the shared TTS streaming response by default, computed by a model-agnostic forced aligner that takes (audio_chunk, text, sr) → list[WordTimestamp].

Suggested scope:

• New shared utility vllm_omni/utils/forced_aligner.py that lazy-loads a forced aligner and exposes align(audio_chunk, text, sr) -> list[WordTimestamp]. Stateless, reusable across all TTS backends.
• Extend the shared streaming response schema in vllm_omni/entrypoints/openai/protocol/audio.py with a word_timestamps field.
• Hook into the shared streaming path vllm_omni/entrypoints/openai/serving_speech_stream.py so every TTS model benefits without per-model changes.
• Offline example under examples/offline_inference/qwen3_tts/ demonstrating the timestamps.
• E2e test in tests/e2e/online_serving/ covering at least Qwen3-TTS and one other model (e.g. VoxCPM2 or CosyVoice3) to prove the cross-model contract.

Marked as good first issue because it's purely additive: no AR / vocoder / scheduler changes — one new utility + one response-schema field + one shared streaming hook.

Alternatives

• Client-side alignment. Caller runs a forced aligner after receiving full audio. Rejected: requires every caller to load a second model and breaks barge-in scenarios that need timestamps before the utterance finishes.
• Per-model forced aligners. Wire it into each TTS model's own streaming code. Rejected: duplicates work across 7+ models and drifts over time. Keeping the aligner model-agnostic is the right boundary.

Additional context

• Shared streaming path: vllm_omni/entrypoints/openai/serving_speech_stream.py, serving_speech.py, protocol/audio.py.
• TTS models that will all benefit: vllm_omni/model_executor/models/{qwen3_tts,voxcpm2,cosyvoice3,fish_speech,omnivoice,moss_tts_nano,voxtral_tts}/.