Multi-Modal Inverse Constrained Reinforcement Learning from a Mixture of Demonstrations