Implement LFU cache to replace LRU cache for speed-up

[qiyunzhu]

Currently Woltka uses Python's built-in functools.lru_cache (least recently used, LRU) to store recent classification results. However, because the frequencies of classification results are evenly distributed across each sample, the alternative policy, least frequently used (LFU) should be a better option. This needs to be implemented and tested manually.