Fix thread safety, implement prompt, fix resource leaks, add history rotation

- Add threading.Lock around asr_pipeline calls (thread-safe inference on shared GPU)
- Implement prompt parameter: propagate through process_file → transcribe → generate_kwargs
- Fix prompt_ids device mismatch (.to(self.device) for CUDA)
- Replace mkdtemp() with create_temp_file() in sources.py (no more leaked /tmp dirs)
- Add max_history_days (default 30) with automatic cleanup of old history directories
- Add intentional comments for SSRF and /config decisions (internal service)
- Lower normalize_audio log level from INFO to DEBUG

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

config.json

@@ -3,6 +3,7 @@
     "model_path": "/home/text-generation/models/whisper/podlodka-turbo",
     "language": "ru",
     "enable_history": true,
+    "max_history_days": 30,
     "chunk_length_s": 28,
     "batch_size": 6,
     "max_new_tokens": 384,