ELEMENTAL: Interactive Learning from Demonstrations and Vision-Language Models for Reward Design in Robotics