Elephant in the Room: Unveiling the Impact of Reward Model Quality in Alignment