On Pitfalls of $\textit{RemOve-And-Retrain}$: Data Processing Inequality Perspective