Handle corrupt instances

[markdryan]

We need some sort of strategy for handling corrupt instances. For example, launcher might detect on start up that the state it maintains about an instance has become corrupted, or the docker container associated by with that instance has been deleted.

Currently, if launcher cannot retrieve an instance state it simply ignores it. This means that it is not reported to the upper layers, and so cannot be deleted, and is essentially leaked on the compute node.

Special handling is also required for instances associated with docker containers that get deleted by some out of bound mechanism, i.e., by someone logging into the compute node and doing

sudo docker rm -f