Lillama: Large Language Models Compression via Low-Rank Feature Distillation