Multigrid projection integration - #201

New ns_solver_params_t.pressure_solver field selects the scalar projection's
pressure Poisson solve: default CG (0, backward compatible), standalone
multigrid V-cycles, or MG-preconditioned CG. The standalone mode subtracts
the RHS interior mean first: the projection RHS is only near-compatible and
the true Neumann system stalls on the incompatible component, while the
CG-based presets are insensitive to it.

Non-2^k+1 grids fail projection init with CFD_ERROR_UNSUPPORTED (probe-init
of the MG hierarchy), and projection_optimized/omp/gpu reject any non-default
selection at init: multigrid has no SIMD/OMP/GPU backend and cross-backend
fallbacks are forbidden.

The previous "zero-init backward compat" test ran make_params() twice, so
both configs went through ns_solver_params_default() and only differed by
NS_PRESSURE_SOLVER_DEFAULT vs (t)0 - the same value - making the memcmp
assertions compare identical configurations.

Build three configs from one genuinely zero-initialized struct that differ
only in pressure_solver: a zero-init run must match the explicit-CG run
bitwise and must differ from the explicit-MG run, proving the zero value
selects CG (not the multigrid path). Anchoring on a zero base avoids
default()'s nonzero source amplitudes, which otherwise confound the compare.

The helper now takes a full ns_solver_params_t so callers control exactly
how each field was initialized.

poisson_solver_tuning.c prints L2=-1.0 when a solve does not return
CFD_SUCCESS (max_iter exit), so the Jacobi row's L2=-1.0e+00 is a
"not computed" marker rather than a real negative error. Call this out
in the Expected Output so readers don't misread it as a valid value.

poisson_solver_create() returning NULL made projection_init return
CFD_ERROR_UNSUPPORTED with no last-error message, leaving callers and tests
without a reason for the rejection. Set an explicit message, matching the
other rejection paths in this init.

projection_init probe-initializes a multigrid solver purely to validate the
grid dimensions, but a default init builds the entire coarse-grid hierarchy
on conforming grids - an O(N) allocation and transient memory spike that the
real pressure solve then rebuilds from scratch.

Pass mg_max_levels = 1 so the probe stops at the finest level. Level 0
borrows the caller's buffers and needs no residual buffer, so the probe now
allocates only the context and one level descriptor. The 2^k+1 dimension
check runs before any of that, so the rejection behavior is unchanged.

The INVALID->UNSUPPORTED remap in projection_init assumed the probe's
CFD_ERROR_INVALID could only mean a non-2^k+1 grid. But poisson_solver_init
also returns INVALID for degenerate grids (nx<3, ny<3, or nz==2), so those
were misreported as UNSUPPORTED with a "requires 2^k+1" message.

Screen degenerate grids up front and return INVALID with a fitting message.
The two statuses drive different recovery: UNSUPPORTED tells the caller to
try another pressure solver, which is useless advice for a grid too small
for any solver. After this screen the probe's INVALID is unambiguous.

Adds test_projection_mg_rejects_degenerate_grid_as_invalid covering a 2x2
grid, alongside the existing non-2^k+1 (UNSUPPORTED) case.