Fix inconsistent state when podagent is down
This fixes some cases (but not all) where there can be an inconsistent
state between what the annotations describes and what CNI configuration
have actually been run.
eg: When a node reboot and Pod foo is started, then annotated BEFORE
podagent has started.
Note that with this change, the CNI configuration MUST be idempotent
because they will be re-run whenever podagent restarts.