How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models