Add NaN/Inf safety check to prevent JSON serialization errors

Fix model loading: upgrade transformers to 5.0.0 and handle safetensors metadata issue

Refactor project structure to Unified AI Voice Detection System