CNI can assign IP address that are no longer attached to the instance

[nckturner]

I had an m4.large (2 ENIs at 10 IP addresses each) which was saturated with pods, and I detached one ENI, then scaled down the pods to 0, then deleted the ENI. Then I scaled the pods back up so that the node was saturated, and every pod entered the running state and and was allocated an IP address by the CNI. But, when I look at the ENIs allocated to the node, there is only 1 ENI. That is, the CNI never requested a new ENI, and instead handed out stale IP addresses to the Pods. This results in connection timeouts when packets are routed to the pods that are behind non-existent IP addresses.