A Fine-Grained Analysis of the LoRA Fine-Tuning Landscape with Implications for Data Selection