v0.8.0

RealtimeTTS 0.8.0 adds an optional Higgs Audio streaming client, experimental word-aligned playback callbacks, fragment lookahead for earlier synthesis, and reliability improvements across playback and several engines.

## Highlights

- Added an optional HiggsEngine client for SGLang-Omni Higgs Audio v3 servers.
- Added experimental, opt-in word alignment and `on_word` callbacks.
- Added fragment lookahead using stream2sentence 1.0.4.
- Added PocketTTS language selection and Qwen pause/resume controls.
- Improved Kokoro voice caching, validation, language handling, and cleanup.

## Notes

HiggsEngine passed protocol-level tests but was not validated against a live server. Forced alignment is disabled by default and remains experimental; slow nonblocking alignment can produce late or missing callbacks.

## Fixes

- Empty synthesis results now fail explicitly.
- Improved immediate-stop cleanup for blocked audio backends.
- Prevented stale alignment results from affecting later playback.
- Fixed PocketTTS GPU state restoration.

Requires Python 3.10–3.14.

[Full changes](https://github.com/KoljaB/RealtimeTTS/compare/v0.7.4...v0.8.0)