Closing the Visual Sim-to-Real Gap with Object-Composable NeRFs