Layer-Wise Evolution of Representations in Fine-Tuned Transformers: Insights from Sparse AutoEncoders