Effect of dataset partitioning strategies for evaluating out-of-distribution generalisation for predictive models in biochemistry