Pandora's White-Box: Precise Training Data Detection and Extraction in Large Language Models