This moves the code to look up significance heuristics information like
background frequency and superset size out of
`SignificantTermsAggregatorFactory` and into its own home so that it is
easier to pass around. This will:
1. Make us feel better about ourselves for not passing around the
   factory, which is really *supposed* to be a throw away thing.
2. Abstract the significance lookup logic so we can reuse it for the
   `significant_text` aggregation.
3. Make if very simple to cache the background frequencies which should
   speed up when the agg is a sub-agg. We had done this for numerics
   but not string-shaped significant terms.