XAI-driven Knowledge Distillation of Large Language Models for Efficient Deployment on Low-Resource Devices