On Partial Optimal Transport: Revising the Infeasibility of Sinkhorn and Efficient Gradient Methods