Learning Closed-loop Dough Manipulation Using a Differentiable Reset Module