Safe Reinforcement Learning with Scene Decomposition for Navigating Complex Urban Environments