Caffeine-based remote cache with adaptive admission windows

[Caffeine 2.7.0](https://github.com/ben-manes/caffeine/releases/tag/v2.7.0) introduces a cool feature of adaptive admission window size to balance between recency and frequency.

This issue is about the idea to apply this in Druid and have admission windows on Historical nodes to "prove" that segment results are "worth" to be propagated to remote dedicated cache nodes (be it memcached or Java servers running Caffeine).

This issue exists to track https://github.com/ben-manes/caffeine/issues/303, where I think the main discussion should happen.

Start with the Caffeine 2.7.0 release notes and the linked Caffeine issue 303, since the issue directs the main discussion there. No Druid files or tests are named; done would require an agreed design for adaptive admission windows on Historical nodes and its interaction with remote memcached or Caffeine-based cache servers.

Written by the indexing model from the issue text.