In high-noise single-cell drug screening datasets, simple mean-prediction baselines can achieve competitive performance compared to complex unregularized models