Is your feature request related to a problem? Please describe.

In the initial implementation of pausing followers, #84252, we pass the current set of overloaded followers to each unquiesced replica tick, and compute the set of followers to pause. This could be somewhat expensive.

Describe the solution you'd like

We should measure the overhead and if it is significant, switch to a system where we only recompute the pausing if there are material changes to the set of overloaded stores which have a follower on this replica.

Describe alternatives you've considered

Additional context

Related to #84252

Jira issue: CRDB-17674