Set guard cells in corners with mixed boundary conditions

[dpgrote]

In the evolve loop, whenever fields are updated, the WarpX specific boundary conditions are applied first (e.g. PEC) and then the periodic (in a call to the appropriate FillBoundary routine). However, the FillBoundary routines will only fill guard cells using valid cells. For mixed BCs then, the guard cells in the corners don't get set to the periodic repeat of the side guard cells along the periodic direction. So, since WarpX applies its BCs first, the changes that are made in the side guard cells are not copied to the corners along the periodic direction, leaving them either zero or with incorrect values.

The fix is to call the PEC boundary routines after the FillBoundary routines, moving them out of the EvolveE and EvolveB routines and making direct calls to them.

Fix #3900

Tasks:

• Move ApplyEfieldBoundary for FDTD
• Move ApplyBfieldBoundary for FDTD
• Do the same changes of PSATD
• Fix CI benchmarks as needed. With careful checking, the tests were failing because the development version was not correct, with the code using corner cells that had not been updated.

[WeiqunZhang]

There are amrex functions for filling ghost cells at the corner of periodic/non-periodic boundaries. https://github.com/AMReX-Codes/amrex/blob/68aa69a11750b3f6ac36ec13b26249f0be80d527/Src/Base/AMReX_FabArray.H#LL1098C35-L1098C43 Not sure if that could be an easier fix than the changes in this PR.

There is also a known issue with PEC boundaried. #3900. Fixing that by passing the correct number of ghost cells to SumGuargCells should also fix the corners.

(Admittedly, I have not read the changes in this PR in detail.)

[dpgrote]

@WeiqunZhang FillBoundary is already being called with the number of ghost cells specified. This seems to be the same as calling EnforcePeriodicity with nghost since both are calling FBEP_nowait. One of the reasons I like my method is that it puts the handling of the boundary in one place, rather than scattered in different routines. In fact, I think I'll update this pull request to put the Apply*FieldBoundary in the 'FillBoundary` calls to reduce code duplication.

The fix for #3900 may be the same as is done here.