Add support for learning without an input domain in offline NSAM (or hybrid approach)

[argaman-aloni]

Try to create the trajectories such that they provide the information provided in the domain such as types etc.
This way we are not bounded by the number of actions in the domain.