Learning Transferable Features for Implicit Neural Representations