Early-twentieth-century cold bias in ocean surface temperature observations