The Effect of SGD Batch Size on Autoencoder Learning: Sparsity, Sharpness, and Feature Learning