Enigmata: Scaling Logical Reasoning In LLMs With Synthetic Verifiable Puzzles