Extend and improve demo

[izderadicka]

Disclaimer: I use Claude Code for the changes, but was reviewing the code - so it's not blind vibe coding. So code looks reasonable from my perspective.

Motivation

I was quite impressed that piper can run on onnx runtime in browser with quite good results , however the original demo was bit limited, so I've played with it and modified it. Key changes:

• onnx runtime to run in multithreaded mode - it require COOP and COEP headers - so that's why there is minimal serve.py
• script (GH pages do not provide possibility to customize headers - so then public hosting should be behind CloudFlare proxy for instance)
• Start playback as soon as possible - anyhow inference is done per sentence - so it can be played right away when it is prepared.
• It also means that model is loaded directly after file changed in input, not waiting for Speak button.
• Highlight sentence, that is being played
• Enable to jump in text by clicking on sentence

By no means it's perfect, but it enabled me to play with speech more interactively and have fun with experiments.