Efficient and Accurate Gradients for Neural SDEs