Unveiling Latent Knowledge in Chemistry Language Models through Sparse Autoencoders