Adaptive Learning of Rank-One Models for Efficient Pairwise Sequence Alignment