Improving Respiratory Sound Classification with Architecture-Agnostic Knowledge Distillation from Ensembles