cdc: Rolling panic/OOM handling

Note: this is a *hypothetical* scenario. We want to make sure
that the system is hardened against such situation.

Understand (and possibly improve) changefeed behavior in the face
of a *hypothetical* scenario where changefeed aggregator panics or OOMs.

The goal is to make sure that if such an incident were to occur, will the cluster
still remain operational? Concretely, verify if the current retry configurations
are appropriate of if the backoff needs to be more aggressive.

How will multiple (10? 100? more) changefeeds that
are in a panic/OOM loop in a single cluster impact such cluster?

Jira issue: CRDB-31389