Variable-Agnostic Causal Exploration for Reinforcement Learning