Contextual Fine-to-Coarse Distillation for Coarse-grained Response Selection in Open-Domain Conversations