[WIP] Implement 3D mass matrices for the implicit solver (direct deposition only)

[RemiLehe]

The full mass-matrix machinery used by the particle-suppressed JFNK solver (implicit_evolve.use_mass_matrices_jacobian) was implemented in 1D and 2D, but not in 3D.

This PR adds the 3D implementation for direct deposition, following the 1D/2D style.

Changes

• MassMatricesDeposition.H: implement the full_mass_matrices branch of the 3D direct deposition kernel, using the same component-index convention as ApplyMassMatrices and the symmetry-reduced deposit of the diagonal blocks (only offsets with dx+dy+dz <= 1).
• ImplicitSolver.cpp:
  • add the 3D symmetry fold in FinishMassMatrices, which copies the remaining diagonal-block components from their point-symmetric mirrors;
  • add the missing dir-2 loop in SyncMassMatricesPCAndApplyBCs and allow implicit_evolve.mass_matrices_pc_width in 3D;
  • replace the blanket 3D assert by one requiring direct deposition. The Villasenor 3D mass-matrix kernel is left as follow-up work and now aborts with a clear message.
• ThetaImplicitEM.cpp: fix an operator-precedence bug in the 3D curl2_BC_mask setup (field_dir + 1 % AMREX_SPACEDIM evaluated to field_dir + 1).
• MatrixPC.H: fix a wrong comment on the component-index mapping.

New CI test

test_3d_theta_implicit_jfnk_vandb: a 3D periodic uniform plasma (16^3 cells, 2x2x2 particles
per cell, 2 species, 10 steps), extending the existing 2D Villasenor and Buneman test. It uses the
theta-implicit solver with the mass-matrices Jacobian and the curl-curl MLMG preconditioner with
mass-matrices response, so it needs no PETSc. It runs in about 25 s on 2 ranks. The analysis
checks energy conservation and the Newton and GMRES iteration counts, all of which are sensitive
to mass-matrix indexing errors.

The parameters shared by the 2D and 3D vandb tests are factored into inputs_base_vandb
(included via the FILE mechanism), and the 2D and 3D analysis scripts are merged into a single
analysis_vandb_jfnk.py that reads the dimensionality and algorithm options from
warpx_used_inputs. The 2D test is unchanged (identical warpx_used_inputs dump).

Validation

• Energy is conserved to 2e-15 (relative); Newton converges in 3 iterations per step and GMRES needs 2 iterations per Newton iteration.
• The solution obtained with the mass-matrices Jacobian agrees with the one obtained with the matrix-free (particle-based) Jacobian to better than 6e-13 on all field quantities, i.e. at the Newton solver tolerance.
• The existing 1D/2D implicit tests (planar pinch, vandb, symmetry) still pass their physics analyses.

🤖 Generated with Claude Code