server: use liveness instead of status record during upgrade checks

The cluster upgrade checks use node status records rather than liveness records to determine cluster membership. This is problematic with clusters from before 20.2, since node decommissioning in those versions simply deleted the liveness record (but not the status record), and so upgrades to 21.1 and beyond will fail since it believes these nodes are still cluster members.

This needs to check against the liveness record instead. We'll have to backport this all the way back to 21.1 to allow such clusters to upgrade to recent versions.

https://github.com/cockroachdb/cockroach/blob/75a55015e83f7e5ff3b2c3c7ee5201f5705414a1/pkg/server/auto_upgrade.go#L126-L134

Jira issue: CRDB-27080

Epic CRDB-27235