feat: backend registry, S2-Pro INT4, progressive segmentation, text cleaning ...

- VRAM снижен 15.4->9.6 GB (max_seq_len 32768->4096, KV cache -4.7 GB)
- Создан реестр бэкендов tts/__init__.py с @register() декоратором
- Бэкенды саморегистрируются: dummy, s2, fish_speech, f5_tts, xtts_v2
- server.py упрощён: create_engine() из реестра, нет _BACKEND_MAP
- _sync_synthesize универсален (нет isinstance проверок)
- ref_text добавлен в TTSEngine.synthesize base class
- Удалён bnb 4-bit код из inference.py (INT4 авто-детект по пути)
- cleanup_text_for_tts(): эмодзи, HTML, URL, маркдаун, спецсимволы
- Прогрессивная сегментация (fast_start_initial=12, fast_start_count=3)
- Документация: README, .env.example, docs/05_usage.md под S2-Pro
- AGENTS.md: актуальное состояние проекта
- .gitignore: .hf_home/, outputs/, voices/*.{m4a,mp4,flac,bak}

Document multilingual F5-TTS behavior and fallback options ...

- Tested ru/en/uk/es/de/fr inference timings

- Note accent and pronunciation caveats

- Outline future fallback/fine-tuning paths

Add explicit reference text config and clean default reference ...

- Add DEFAULT_REF_TEXT setting to skip Whisper transcription

- Pass ref_text through F5 backend and warm-up path

- Document new env variables and warm-up usage

- Update latency notes with explicit ref_text timing

Initial voice TTS server with F5-TTS backend, WebSocket streaming and warm-up ...

- FastAPI WebSocket API for real-time chunked TTS

- F5-TTS GPU backend with reference caching and auto-transcription

- Dummy backend for offline tests

- Streaming text segmenter, session state, stop/resume

- Startup warm-up, PCM16/base64 audio, basic tests

- Documentation: architecture, protocol, usage, roadmap, technical notes