Text this: Benchmarking ML in ADMET predictions: the practical impact of feature representations in ligand-based models