Cooperative Policy Learning with Pre-trained Heterogeneous Observation Representations