Fix energy conservation in 1D VlasovAmpere runs with diagonal scaling

[max-models]

Summary

Since #739 made diagonal_scaling=True the default of MassMatrixPreconditioner, the total energy of 1D VlasovAmpereOneSpecies runs (one element in e2 and e3) is no longer conserved: it drifts by ~1.5e-4 per step instead of ~1e-12. This breaks test_verif_VlasovAmpereOneSpecies.py::test_energy_conservation in #738. This PR fixes two bugs:

1. VlasovAmpereCoupling: update the ghost regions of the mid-point field before the push. The pusher kernel reads the ghost regions of e_sum directly. They were only correct as long as the Schur solver happened to return valid ghost regions. With the diagonal scaling it doesn't, because its last step is a StencilDiagonalMatrix.dot, which writes only the owned entries. In 1D runs every evaluation reads ghost values. The other couplings (EfieldWeightsCoupling, CurrentCoupling6DCurrent, CurrentCoupling5DCurlb) already update the ghost regions before pushing.

2. Diagonal scaling: include the band entries that wrap onto the diagonal. In a periodic direction with fewer points than the band width (2p + 1 > n, e.g. a single element), the band entries at offsets k = m n also lie on the main diagonal. StencilMatrix.diagonal() returns only the entry at k = 0, so the diagonal D of the mass matrix was wrong, and MassMatrixPreconditioner with diagonal scaling was no longer exact on the logical cube. Fixed with _stencil_local_diagonal in _local_diagonal, which MassMatrixPreconditioner._core_diagonal now uses.

Tests

• New test_mass_matrices.py::test_mass_preconditioner_diagonal_scaling_periodic_exact: checks that the preconditioner with diagonal scaling inverts M0 to M3 exactly on the Cuboid, including grids with one element. Without the fix, 6 of the 12 cases fail.
• Local runs, serial and on 2 ranks: test_preconditioner_transpose.py, test_stiffness_preconditioner.py, test_mass_matrices.py::test_mass_preconditioner* and test_poisson.py (127 / 151 passed). With Fix total energy at t=0 in Vlasov-Ampère/Maxwell models (ghost regions, control variate) #738 merged in, test_verif_VlasovAmpereOneSpecies.py passes (3 passed).

🤖 Generated with Claude Code

[spossann]

No changes here, got probably fixed in #738