Understanding Neural Coding on Latent Manifolds by Sharing Features and Dividing Ensembles