On Principal Component Regression in a High-dimensional Error-in-variables Setting