docs(qwen3-tts): add real-time streaming tuning guide

[akaiHuang]

Summary

Add a practical real-time tuning section for Qwen3-TTS streaming in the model docs.

What this adds

• Explains the streaming_interval = chunk_tokens / 12.5 relation for Qwen3-TTS codec rate.
• Adds a small tuning matrix (chunk_tokens 2/4/6/8) with trade-offs for latency vs smoothness.
• Recommends stream=True, streaming_interval=0.32 as a pragmatic default for real-time usage, with guidance to move to 0.48 when smoother chunks are needed.

Why

Users deploying live translation / assistant pipelines often need concrete first-pass values rather than only API descriptions. This doc update shortens trial-and-error and aligns expectations around chunk boundaries vs latency.

Scope

Docs-only change. No runtime behavior changes.

[akaiHuang]

Hi @Prince_Canuma @Blaizzy,

I've added a practical real-time streaming tuning guide for Qwen3-TTS in the documentation.

This is based on extensive benchmarking on M1 Max (4-bit). It shows:

• How to tune chunk_tokens / streaming_interval
• The trade-off between first-chunk latency and audio smoothness
• Recommended settings for real-time use cases (live translation, voice agents, P2P translator, etc.)

The guide turns the previous ~4.31s first-chunk into a tunable 0.317s–0.546s range.

Would love your feedback on whether this is useful and if any adjustments are needed before marking it ready for review.

Thanks!