Bridging the Domain Gaps in Context Representations for k-Nearest Neighbor Neural Machine Translation