Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles