Interpreting and Steering Protein Language Models through Sparse Autoencoders