improve MPI scaling of set_maxwell_dielectric

[stevengj]

Right now, MPB calls back to the meep_mpb_eps Meep for every pixel of the mode profile to find the ε profile for the eigenmode solver, and this performs a collective communication. This is a problem for scaling, because, you then have huge numbers of small collective calls, one per pixel in the eigenmode region.

Should be straightforward to fix. Probably something like:

• Call set_maxwell_dielectric with a dummy function (just returns 1 everywhere), which caches the inv(ε) value for owned chunks in an array and zero for nonowned chunks. (i.e. pass parallel=false to get_chi1inv).

• Call sum_to_all once on this cache array to synchronize the values.

• Call set_maxwell_dielectric again, this time returned the cached values from each process.

Hence, only a single collective call is required in the second step.