Diffusing States and Matching Scores: A New Framework for Imitation Learning