Layer-Wise Perturbations via Sparse Autoencoders for Adversarial Text Generation