Improve broadcast message cache management

• Follow up on comment gitlab-com/migration#449 (comment 74221210):

Ohhh I got it. During failover, the secondary becomes a primary, and the cache key expiration becomes indefinite again.

So this only affects instances that are demoted from primary to secondary. 😢

I'll create an issue for that in EE, which I feel could be done post migration.

On a related note, should we add a migration to clear this key in any case?

That would be fine, although due to the above, I think we may as well resolve it at the same time in the caching logic. I.e. Maybe there is a good spot to delete the cache key that would resolve all of this.

• Should we remove the 10.8 upgrade note to clear cache, since it should be handled by this issue? https://docs.gitlab.com/ee/administration/geo/replication/updating_the_geo_nodes.html#upgrading-to-gitlab-10-8