Avoid switchover after migrating a single-pod cluster

[Shubham-Padkonde]

Problem description

Migrating a single-pod cluster off an unschedulable node recreates the master, then attempts a Patroni switchover to itself using the deleted pod's IP. masterCandidatePod starts as oldMaster, making the existing recreate-only branch unreachable.

Leave the candidate unset until a replica is selected. Single-pod clusters now take the existing recreate-and-return path; clusters with replicas still select a candidate and switch over to it.

Added tests exercise successful single-pod relocation without any Patroni request, deletion failure, relocation to the same node, and successful switchover to a healthy replica. The migration parameter documentation now explains the single-pod behavior and downtime.

Validation on Linux with Go 1.26.4:

• Before the fix, the new relocation regression fails on the unexpected Patroni request; after the fix, all migration and candidate-selection cases pass.
• make test passes across the repository.
• go test -race ./pkg/cluster -run 'TestMigrate|TestGetSwitchoverCandidate' -count=1 passes.
• make local builds the operator.
• gofmt and git diff --check pass.

The tests use the Kubernetes fake client, pod events, and a mocked Patroni HTTP client. Live Kubernetes node rotation and database end-to-end tests were not run.

AI assistance: Codex (GPT-6) assisted with investigation, code, tests, documentation and this description.

Linked issues

Fixes #3059. Proposed approach posted in the issue before implementation.

Checklist

• Go code is formatted.
• Generated API code checked: no API fields changed, so no regeneration is needed.
• Configuration/CRD/chart impact checked: no configuration options added.
• Behavior is covered by unit tests.
• Existing open PRs checked for overlap; none found for Operator cache obsolete IP address during cluster migration #3059.

[Shubham-Padkonde]

The COMPR/pr-labels check requires one of the project classification labels. Could a maintainer add bugfix? This account cannot apply labels. The fix addresses #3059; full Go tests, focused race checks and the operator build passed locally, as detailed above.