Unauthenticated code execution via unsafe model loading

Unsafe deserialization in Applio 3.2.8-bugfix and earlier allows remote attackers to execute code through a crafted model file. The inference and text-to-speech interfaces pass a user-controlled model path through change_choices to get_speakers_id, which calls torch.load without restricting loading to tensor weights. No authentication or user interaction is required when the affected interface is network-reachable, and successful exploitation runs code with the Applio process's privileges.