[sumimakito]

This pull request adds streaming Sortformer diarization to Sherpaw:

• Label up to eight speakers every 10 ms with NVIDIA Nemotron 3 Diarization on WebGPU, from @sherpaw/speaker-diarization/sortformer. It also marks when two people talk at once.
• Choose 1.04 s or 0.64 s latency. A benchmark on the device recommends the one that keeps up with live audio: on an Apple M4, the model runs about 2.7 times and 1.9 times faster than real time.
• Turn the probabilities into speaker turns. A turn starts where the new speaker becomes more likely than the previous one, so the first word of a quick reply no longer joins the previous turn.
• Download the model with a script that stores its weights as int8: 104 MB instead of about 400 MB, with nearly the same output.
• In the demo, a Nemotron page shows the turns live and how far each step is behind real time.

Next: publishing the int8 model on Hugging Face, so that the deployed demo can load it.

[github-actions]

✅ Deploy to Cloudflare Workers (Preview) for sandbox ready!

Name	Link
🔍 Latest deploy log	https://github.com/moeru-ai/sherpaw/actions/runs/38079427263
😎 Deploy Preview	https://02b04a4f-moeru-ai-sherpaw.moeru-ai.workers.dev
🚀 Pull Request Preview	https://pr-31-moeru-ai-sherpaw.moeru-ai.workers.dev