More accurate gradient with objective functions containing Fourier arguments

[smartalecH]

As usual, meep interpolates the DFT fields from the yee grid. This interpolation introduces an error in the adjoint gradient calculation unless we backpropagate through it (just multiply by the transpose of the interpolation). We haven't yet implemented this, but probably should (when we have time and need the feature).

Ironically, the eigenmode arguments, which are much more complicated (and technically a function of the Fourier fields...) don't have this issue because we already took this into account (see #1270).

[smartalecH]

#1959 seems to have taken care of this.