Incorrect gradients from fixed-point differentiation of eigh-based C4v CTMRG using `:diffgauge` iteration scheme

[leburgel]

As mentioned in #330, computing PEPS gradients through fixed-point differentiation of eigh-based C4v CTMRG using the :diffgauge iteration scheme in which we compute the gradient using the pullback of a single gauge-fixed iteration gives rise to incorrect results. While the gauge fixing of the environment itself definitely works, the issue might be caused by a subtlety in the gauge fixing of the eigenvalue decomposition (potentially related to #331).

[leburgel]

This problem doesn't seem to occur for QR CTMRG (tested in #329), while it uses the same gauging routine. So it seems like a problem specifically with gauge fixing the eigenvalue decomposition.