fix: add cleanup of published ResourceSlices on shutdown

[ttsuuubasa]

Summary

Ensure the driver deletes the ResourceSlices it published when it is shut down. Previously, on SIGTERM the driver simply stopped without removing its slices, leaving stale ResourceSlice objects in the cluster.

Motivation

The driver currently leaves the ResourceSlice objects it published in the cluster even after it is shut down. This is especially problematic when switching between dynamic scaling (DDS + CDI DRA) and manual scaling (CRO only): leftover slices from CDI DRA interfere with the manual mode, forcing operators to identify and delete those slices by hand.

Reclaiming slices automatically on CDI DRA shutdown removes that manual step and reduces operator burden. It also improves the uninstall experience.

This change targets the normal shutdown path (e.g., scaling the Deployment from 1 to 0, or a regular pod termination). Failure-mode shutdowns (crash, OOMKill, node failure, etc.) are out of scope — orphaned slices in those cases are accepted as an acknowledged limitation.

Changes

• pkg/manager/manager.go:
  • Add cleanupResourceSlices invoked via defer in StartCDIManager.
  • It calls Controller.Stop() for each driver, then lists the cluster-scoped slices for that driver (FieldSelector: spec.driver=<name>,spec.nodeName=) and deletes only those whose pool name matches a pool we published. A fresh 30s context is used since the parent context is already canceled at this point.
• main.go:
  • On SIGTERM, explicitly cancel() and then wait on errChan so the deferred cleanup in StartCDIManager completes before the process exits.
  • Without this, the goroutine and the main routine race and the cleanup may be killed mid-flight.

Design notes

• Controller.Stop() is called explicitly even though ctx cancellation alone would stop the goroutines, because only Stop() blocks via wg.Wait() and thus prevents a race where an in-flight publish recreates a slice we just deleted.
• Cleanup deletes slices one-by-one after filtering by published pool name, rather than using DeleteCollection with spec.driver=<name>. This guards against deleting slices owned by another driver that happens to share the same driver name.