Discovering High-Quality Chess Puzzles with Offline Reinforcement Learning