Understanding and Mitigating Membership Inference Risks of Neural Ordinary Differential Equations