Prior to NumPy 1.21 the low and high values could be reversed, and
samples were calculated using `low + (high - low) * random()`. This was
changed in NumPy 1.21 so that if `low > high` an exception is raised;
see https://github.com/numpy/numpy/pull/17921 for details.

We can preserve the original behaviour by negating low and high and the
samples values. This means we obtain the same samples for NumPy < 1.21
and NumPy >= 1.21.