Work needs to be done before declaring wanting to transit between
running and done states. This work is, asking the underlying transport
to kill its instances and clean stuff on the remote computers. Thus,
declaring that the work is done is wrong. The work is done when this
step is done.

To enable this asynchonous possibility a task can now ask the execution
of a callback function when a changing state is possible. It's a way to
execute an out-of-workflow operation still inside the state-machine
transitions.