Consistency checks currently run asynchronously with a background context on all range replicas. If the caller gives up on these (e.g. because one of them errors), we don't cancel the others. As the caller goes on to run further consistency checks, this can cause a build-up of abandoned consistency checks. We should fix this such that these checks are cancelled when no longer needed.

See draft PR in #76855.

Jira issue: CRDB-13588