Cross-seed explainability using Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoders