LoRA vs. Full Fine-Tuning: An Illusion of Equivalence