Beyond the Hard Budget: Sparsity Regularizers for More Interpretable Top-k Sparse Autoencoders — ThinkLLM