Sparse Autoencoders Do Not Find Canonical Units of Analysis