Switch FMM formulations to new implementation

[mfschubert]

This PR is a further step to addressing https://github.com/facebookresearch/fmmax/issues/55

• Replace the old pol, jones, jones direct and normal implementations with new implementations that avoid iterative field optimization and instead use a single Newton step to find the optimal vector field.
• Adjust vector field objective so that it depends upon the number of terms in the Fourier expansion, such that the field resulting from a 2x2 supercell with 4x the number of Fourier terms is identical to that of the base unit cell.

The hyperparameters (smoothness loss weight and fourier loss weight) were adjusted to achieve good convergence of simulated quantities with the number of Fourier terms in the expansion. For the uLED, this new implementation gives essentially unchanged results when compared to the earlier implementation (see figure).

The old implementations still remain in the code; these should be removed in a follow-up PR.

[smartalecH]

Does this resolve any of the power conservation issues you were seeing?

[mfschubert]

Does this resolve any of the power conservation issues you were seeing?

No it is unrelated. For issues related to power conservation I expect that improved s-matrix treatment will be needed.