Better handling of different servers having different versions of broadcast table segments

### Description
Broadcast data sources can potentially have different segment versions on different servers. It can result in some inconsistencies
- If a query involving broadcast data source is evaluated on different servers (such as join), the end result might be incorrect since different servers could have different versions of data.
- Broker may cache the result of join queries on its end. During the query execution, historicals may have an older version of data. However, once the broker caches it, the historicals will not be re-queried even if they update their data with a newer version of data. Thus, brokers will be left with stale query results. The scenario is described here - https://github.com/apache/druid/pull/10366/files#r494684538

Start by reading the broadcast data source, broker, historical, and join-query scenarios described here, then review the linked discussion in PR #10366. Define how differing segment versions should be handled across servers and broker caching, with completion requiring consistent query results and no stale cached joins after historicals update.

Written by the indexing model from the issue text.