Separate planning commands from running them?

[sourcefrog]

As the logic for what commands to run gets more complex it might be useful to separate the logic that plans commands to run, from the code that actually runs them.

Then we could write some tests that only inspect the generated commands without actually running them, which would make the test suite faster. However the speed would come at the risk of the tests being perhaps less realistic: we wouldn't catch cases where the cargo command semantics don't do what we expect.

Also, this might just make the code easier to follow.