It seems that with the hierarchical algorithm we end up comparing the
same things several times. Due to their low level in the hierarchy, and
comparatively large size, MIME data comparisons are especially affected
by this. For this reason, we simply add a LRU cache on the comparison.

We can only cache string type mime data comparisons (hashable). Also
added a cache on compare_text_approximate, as profiling showed that a
hash there gives a reasonable amount of hits.