End-to-end learning of multiple sequence alignments with differentiable Smith–Waterman