feat(radio): replace robot TTS with Piper Neural voices & add instant voice preview

- Integrate local Piper ONNX neural TTS engine (Dmitri, Irina, Denis, Ruslan)
- Add GET /api/radio/preview-voice endpoint with cached pre-rendered samples
- Add '▶ Прослушать голос' button in Web UI for 1-click preview before generating
- Add automated filter to strip LLM refusals and fall back to clean radio phrasing

app/radio/manager.py

@@ -146,11 +146,8 @@ class RadioManager:
             "playlist_ready": playlist_ready,
             "speech_clips_count": len(speech_files),
             "speech_clips": speech_files,
-            "default_voice": settings.RADIO_VOICE,
-            "available_voices": [
-                {"id": "ru-RU-DmitryNeural", "name": "Дмитрий (Харизматичный DJ)", "gender": "Мужской"},
-                {"id": "ru-RU-SvetlanaNeural", "name": "Светлана (Мягкий вечерний голос)", "gender": "Женский"}
-            ]
+            "default_voice": settings.RADIO_VOICE or "ru_RU-dmitri-medium",
+            "available_voices": dj_host.get_available_voices()
         }
 
 radio_manager = RadioManager()