Prompt-Induced Linguistic Fingerprints for LLM-Generated Fake News Detection