Bug fix

[returnddd]

I think I have found a bug in _log_importance_weight_matrix method.

Striding (M+1) is just batch_size, which jumps rows while the column index is same (vertical move), but I think the intended implementation is a diagonal move on the matrix.