Improving Neural ODEs via Knowledge Distillation