Initially just support --image to change the image.

I was trying to preserve the ordering, but we sort all the args
anyway. Using sets and subtracting them performs much better than
list comprehensions.

Make the tests use nose instead of unittest, and use properly
parameterized arguments. This makes it a lot easier to see
what fails.

To get commands, env vars etc, we need the image's defaults to
do a diff.

I knew I'd forgot something.