High-Entropy Token Selection in Reinforcement Learning with Verifiable Rewards (RLVR) Improves Accuracy and Reduces Training Cost for LLMs