Accuracy of calculations depend strongly on rotational orientation

[kvhuguenin]

In the test called test_random_structure in test_values_ewald.py, the agreement between the target energy (from GROMACS) becomes worse by 1 order of magnitude if the random rotation that we apply is changed to
rot_2 = rot_z(0.123453) @ rot_x(1.2452)
instead of the current value of
rot_2 = rot_z(1.23456) @ rot_x(0.82321)

This big change in accuracy affects not only the PME calculator (which is not rotation invariant due to the mesh), but also the Ewald calculator which should in principle be invariant if the k-space sum is well-converged.