Learning to Distill Global Representation for Sparse-View CT