Fix bug for wave ports

[sebastiangrimberg]

• Fix wave ports bug where the quadrature order was not being set correctly, and thus the eigenvalue solve would fail (especially for high-order).
• Revisited wave port formulation, making a small update to use the variational formulation from https://www.jstor.org/stable/4099985 (see also https://arxiv.org/abs/2302.11994, both replace https://ieeexplore.ieee.org/iel1/22/2793/00085399.pdf) which improves the eigenvalue problem by avoiding the introduction of many modes at \beta = 0.

Note: The re-baselined regression tests for cpw/cpw_wave_uniform and cpw/cpw_wave_adaptive are expected: The quadrature rule fix changes the computed mode shape slightly which affects both the propagation constant as well as the excitation of the model in the driven solve. The changes (all seem to be on the order of a few percent) seem justified in this regard.

[hughcars]

I see Dtt on the previous diff and Dnn now, but am not seeing it in the block system above. From here it looks like an empty matrix with a diagonal of zeros, is this just to ensure some memory in the system matrix for applying the essential bc elimination? I assume then this is the reason for the difference compared to the nullptr initialization for the off diagonal blocks of B in GetSystemMatrixB.

[sebastiangrimberg]

That's right. HypreParMatrixFromBlocks does not allow for a fully empty row or column so that it can generate the right block sizes from the input. So, we just explicitly pass a zero diagonal block for Bnn.