Fix LINELET with mpi, add regression case for laminar NACA0012

[Syphonicc]

Addresses the regression-test part of #2877.

Adds a serial regression case for LINEAR_SOLVER_PREC= LINELET. There is currently no LINELET coverage in TestCases.

TestCases/navierstokes/naca0012/lam_NACA0012_LINELET.cfg is the laminar NACA0012 config with one added line, LINEAR_SOLVER_PREC= LINELET. The baseline uses the default ILU(0).

Note: it is based on the updated lam_NACA0012.cfg from #2866 (ITER instead of the deprecated EXT_ITER, CFL 30), so on current develop the two files differ in more than that one line until #2866 is merged. This PR does not depend on #2866; the new file parses on develop as-is. The registration is placed after cylinder_lowmach, away from the lam_naca0012 entry added in #2866, so the two PRs don't touch the same lines.

At test_iter = 20 the two cases already differ, so the test is sensitive to the preconditioner:

Preconditioner	rms[Rho]	rms[RhoU]	rms[RhoV]	rms[RhoE]
ILU(0) (lam_NACA0012)	-5.522140	-3.127430	-3.286926	-0.109971
LINELET	-5.282867	-2.845208	-2.852421	0.154542

test_vals are from a local serial clean build of current develop (8118ee0, GCC 15.2.1). The same procedure reproduces the #2866 lam_naca0012 values exactly. I'll update them if CI differs.

This does not address the segfault reports in #2877, and only covers the serial path.

[bigfooted]

can you check parallel?

[Syphonicc]

Good catch LINELET segfaulted in parallel with ≥4 ranks (1 and 2 ranks ran fine, which is why it didn't show up before).

Cause

• Partitions without wall markers have no linelets, so CSysMatrix::BuildLineletPreconditioner never allocated the per-thread working vectors but still indexed them → segfault (CSysMatrix.cpp, confirmed with gdb on the rank that had no walls).
• CGeometry::GetLineletInfo used linelets.empty() as its "already computed" check. On ranks with no linelets it recomputed on every call, including the MPI_Allreduce/MPI_Allgather, while the other ranks returned early → mismatched collectives.

Fix

• Always allocate the linelet working vectors (they're small, MAX_LINELET_POINTS per thread).
• Added an explicit computed flag to CLineletInfo so every rank computes it exactly once.
• Skip the "points in each line" message when there are no linelets globally (avoids a divide by zero).

Tests

• Added lam_naca0012_linelet to parallel_regression.py with mpirun -n 4. The default 2 ranks doesn't trigger the bug, because both partitions touch the airfoil.
• Merged develop and updated the serial test_vals.
• Checked the case runs at 1, 2, 4 and 8 ranks.

Fixes #2877.

[rois1995]

LGTM, thanks! Tested: on develop this case crashes on 4 and 8 ranks; with this PR it runs on 2, 4 and 8.

[Syphonicc]

Thanks for testing! CI hasn't actually run yet: all workflow runs on this PR show action_required, so they need a maintainer to approve them. On slots: the CI image uses Open MPI, which counts physical cores by default, so I've added --oversubscribe to the 4-rank command to be safe (the harness already relies on the Open MPI-specific --allow-run-as-root). Header updated as well.

[Syphonicc]

CI findings:

ASAN (serial): a real, pre-existing bug that this PR's serial LINELET test exposed. In GetLineletInfo, nColors is su2uint (uint32_t) but is sent with MPI_UNSIGNED_LONG in the Allgather, so 4 bytes past the variable are read (stack-buffer-overflow in the no-MPI wrapper, and garbage colour offsets with real MPI). Fixed by declaring it unsigned long. Verified with an ASAN build: clean, and the test values are unchanged.

ramp_msw (parallel): looks unrelated. It uses ILU, so LINELET is never called, and it passes locally with residual deltas of about 0.09 against its 0.2 tolerance. It seems to be a noisy test.

[pragyaangaur]

Nice fix. Testing at 4 ranks was the key insight, since at 2 ranks every partition touches the airfoil and hides the crash. One suggestion for a follow-up: the test only covers the fix if the partitioner leaves a rank with no wall nodes, so a small assertion or log of the zero-linelet rank count would keep it from losing that coverage if the partitioning changes.