Minimizing the Age-of-Critical-Information: An Imitation Learning-based Scheduling Approach Under Partial Observations