Learning Multi-Level Features with Matryoshka Sparse Autoencoders