Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders