Understanding Likelihood Over-optimisation in Direct Alignment Algorithms