Learning Bilinear Models of Actuated Koopman Generators from Partially-Observed Trajectories