B-field computation more stable than H-field computation

[OrtnerMichael]

This test shows that B-field computation is more stable than H-field computation.
Don't really understand why - probably something to do with the digits, but cant figure this out in floating point.

# Dipole <> Sphere
def test_core_physics_dipole_sphere():
    """
    dipole and sphere field must be similar outside of sphere
    moment = magnetization * volume
    near field tests
    """
    obs = np.array([(1, 2, 3), (-2, -2, -2), (3, 5, -4), (5, 4, 0.1)])
    dia = np.array([2, 3, 0.1, 3.3])
    pol = np.array([(1, 2, 3), (0, 0, 1), (-1, -2, 0), (1, -1, 0.1)])
    mom = np.array([4 * (d / 2) ** 3 * np.pi / 3 * p / MU0 for d, p in zip(dia, pol)])

    B1 = magpy.core.magnet_sphere_field(
        field="B",
        observers=obs,
        diameter=dia,
        polarization=pol,
    )
    B2 = magpy.core.dipole_field(
        field="B",
        observers=obs,
        moment=mom,
    )
    np.testing.assert_allclose(B1, B2, rtol=0, atol=1e-16)

    H1 = magpy.core.magnet_sphere_field(
        field="H",
        observers=obs,
        diameter=dia,
        polarization=pol,
    )
    H2 = magpy.core.dipole_field(
        field="H",
        observers=obs,
        moment=mom,
    )
    np.testing.assert_allclose(H1, H2, rtol=0, atol=1e-11)