Use complex-aware preconditioner in BoundaryMode solver

[simlapointe]

Improve BoundaryMode linear solver convergence when lossy boundary conditions are present by including complex terms in the real-valued preconditioner ($P = A_r + A_i$, like in the 3D case) or using an exact complex-valued preconditioner when "WavePorts.ComplexCoarseSolve" is true. Closes #920.

[hughcars]

Confirmed the block assembly has to happen before the non-port ranks return: HypreParMatrixFromBlocks allgathers on the FE-space comm. benvial's numbers on #920 are the evidence this works. Nits inline, then rebase onto main once #919 is in and merge.

[hughcars]

Suggested change

	// the FE space communicator. Wave-port sparse-direct solvers run on a subcommunicator,
	// so assemble the complex-aware preconditioner on all ranks before non-port ranks
	// return. The resulting real operator is already in the form expected by the wrapped
	// sparse solver, avoiding another collective assembly in MfemWrapperSolver.
	// the FE space communicator. Wave-port sparse-direct solvers run on a subcommunicator,
	// so assemble the complex-aware preconditioner on all ranks before non-port ranks
	// return. direct_pc_op only carries the assembled real matrix (2N x 2N in the
	// doubled-real case) into MfemWrapperSolver::SetOperator; its Mult is not meaningful.

[hughcars]

With global "PCMatReal": true, port_linear.pc_mat_real is inherited and the !linear.pc_mat_real guard in ModeEigenSolver::Solve skips this entirely, so an explicit per-port "ComplexCoarseSolve": true is silently ignored. Give the explicit override precedence:

Suggested change

	port_linear.complex_coarse_solve = *data.complex_coarse_solve;
	port_linear.complex_coarse_solve = *data.complex_coarse_solve;
	port_linear.pc_mat_real = false;

Note this also means an explicit per-port false gives Ar + Ai rather than Ar. I think that is the right reading of a per-port choice, but if you would rather preserve PCMatReal for that case, only clear it when the override is true.

[hughcars]

The MaxIts/KSPTol fix is a separate bug, give it its own bullet. This will also need to move to In progress when you rebase.