From Text to Trajectory: Exploring Complex Constraint Representation and Decomposition in Safe Reinforcement Learning