Safety-aware Causal Representation for Trustworthy Reinforcement Learning in Autonomous Driving