GMTS: Gradient Magnitude-based Token Selection Improves RLVR Training for LLM Reasoning | TickerVault