Can sparse autoencoders be used to decompose and interpret steering vectors?