Slight precision error in `test_polydiff` on Darwin systems

[loicreynier]

As reported by @Aleksanaa in NixOS/nixpkgs#426180, there is a slight precision error on Darwin (macOS) systems in test_polydiff probably due to a different BLAS backend:

        nodes = dmat.nodes
        assert np.allclose(nodes[0], dmat.xmin)
        assert np.allclose(nodes[-1], dmat.xmax)
        func = nodes**2
        dfunc = 2 * nodes
        d2func = 2.0
        d1_cheb = dmat.at_order(1) @ func
        d2_cheb = dmat.at_order(2) @ func
        d3_cheb = dmat.at_order(3) @ func
        assert np.allclose(d1_cheb, dfunc)
        assert np.allclose(d2_cheb, d2func)
>       assert np.allclose(d3_cheb, 0.0, atol=1e-6)
E       assert False
E        +  where False = (array([-1.27562089e-07, -3.15194484e-08,  3.28873284e-09,  8.14907253e-10,\n       -2.24190444e-10, -3.31965566e-11,  6...9,  4.74392436e-09,\n       -5.41331246e-09,  1.37370080e-08, -8.38190317e-09, -9.08970833e-07,\n       -1.10268593e-06]), 0.0, atol=1e-06)
E        +    where  = np.allclose

tests/test_polydiff.py:49: AssertionError

=========================== short test summary info ============================

FAILED tests/test_polydiff.py::test_cheb_scaled - assert False

========================= 1 failed, 7 passed in 0.52s ==========================

This is not a big deal and I cannot test it on my system but this could probably be fixed using something like

# Platform-aware tolerance
atol = 1e-6
if platform.system() == "Darwin":
	atol = ...

# ...
	assert np.allclose(d3_cheb, 0.0, atol=atol)