With the introduction of the read committed isolation level, transactions running under the isolation level will establish a new read snapshot on each statement boundary (#100133). The fact that each statement runs under a different read snapshot also implies that a savepoint rollback should be able to recover from a non-aborting transaction retry error.

This is currently supported for serializable transactions, but only for savepoints captured at the very beginning of the transaction:

We will need to generalize the mechanism and decouple it from transaction epochs.

This is a foundational change needed to support #100145.

Jira issue: CRDB-28421