Given a parameter tensor param, its gradient grad, and a learning rate lr, return the parameter after one plain gradient-descent update step: param - lr * grad.