Resolve singularities for matched thru and line in determine_reflect

[eendebakpt]

Fixes #865

This PR allows the determine_reflect (used in TRL calibration) to run a matched thru and line. The approach is inspired by the comment #865 (comment) from @Ttl

[Ttl]

I tested the changes and it seems to work fine. No issues with the functionality, but few code quality comments.

[eendebakpt]

@Ttl When making the review changes I found a problematic case. I could solve the case by increasing epsilon in the regularization from 1e-7 to 1e-6, but I am not familiar enough with the material to judge whether 1e-6 is too large.

I changed the root selection criterium to one from the original TRL paper. It works better for the cases I tested, but perhaps we need some more test cases.

On current master the case from the regression test results in NaN values in the outcome.

[Ttl]

This chooses the root based on which one gives lossy line. However it is inaccurate for very low loss line such as waveguide. The original one is also described in the paper and is more robust. Keep the old root choice algorithm.

[eendebakpt]

@Ttl I reverted the root choice, but now I have a failing test which I do not know how to fix. The failing test is:

        freq = rf.F(434615384.6153846e-9, .7, 1, unit = 'GHz')
        medium = Coaxial.from_attenuation_VF(freq, att = 3.0, VF = .69)
        thru= medium.line(0, 'm')
        line = medium.line(0.12, 'm')
        short = rf.Network(f=freq.f, s=[[[(-1.+0.017870117714376983j)] ] ])

        r = determine_reflect(thru, rf.two_port_reflect(short, short), line)
        npy.testing.assert_array_almost_equal( r.s, short.s)

It does pass with the rootChoice = npy.abs(x1) < npy.abs(x2) test, but not with the rootChoice = abs(x1 - e2) < abs(x2 - e2). Do you have any suggestions on how to address this? (maybe this is a case that we should not expect to work, but it seems like a reasonable case to me)

[Ttl]

This passes on current master:

    def test_determine_reflect_regression(self):
        freq = rf.F(434615384.6153846e-9, .7, 1, unit = 'GHz')
        medium = Coaxial.from_attenuation_VF(freq, att = 3.0, VF = .69)
        thru = medium.line(0, 'm')
        line = medium.line(0.12, 'm')
        short = medium.load(-1.+0.017870117714376983j)

        # workaround division by zero
        thru.s[:,0,0] += 1e-7
        thru.s[:,1,1] += 1e-7
        r = determine_reflect(thru, rf.two_port_reflect(short, short), line)
        npy.testing.assert_array_almost_equal(r.s, short.s)

[eendebakpt]

@Ttl Ok. I tried before only to regularize t, but with the regularization on s now all tests are passing.