Support persistent cache to speed up duplicated inference

[yuancu]

When I experiment with different kinds of retrieval methods, it always takes more than half an hour to retrieve subgraphs for 100 samples. The in-memory cache get lost when I re-run the script.

I decide to adopt cachier to speed up retrieval.