Accuracy on the wrong line: On the pitfalls of noisy data for out-of-distribution generalisation