The Representation Landscape of Few-Shot Learning and Fine-Tuning in Large Language Models