On Model Identification and Out-of-Sample Prediction of Principal Component Regression: Applications to Synthetic Controls