Add GPU plain SOR (Block SOR) linear solver, closing last backend gap

[shaia]

Plain SOR is sequential (each cell depends on already-updated left/down neighbors), so a naive GPU mapping fails. The GPU variant uses red-black tile coloring: each CUDA thread sweeps an 8x8 interior tile in row-major order (Gauss-Seidel inside the tile), and tiles are colored so no two simultaneously active tiles share a halo boundary. Two kernel launches per iteration (color=0 then color=1) serve as the synchronization barrier, mirroring the existing Red-Black SOR GPU solver.

The double-buffered tile design tried first diverged at optimal-omega (~1.85 on 33x33) because Jacobi-coupled tiles with omega>1 are unstable. Red-black tile coloring is a consistent ordering and is provably convergent for 0<omega<2, so auto optimal-omega works correctly.

Files changed:

• poisson_gpu_primitives.cuh: new lin_gpu_kernel_block_sor_tile_sweep kernel
• poisson_solver_sor_gpu.cu: new solver (ctx, init, destroy, solve, factory)
• linear_solver_internal.h: factory declaration under #ifdef CFD_HAS_CUDA
• linear_solver.c: dispatcher case for POISSON_BACKEND_GPU / POISSON_METHOD_SOR
• lib/CMakeLists.txt: add .cu to CFD_CUDA_SOURCES
• CMakeLists.txt: test executable, link, and ctest registration
• tests/math/test_poisson_sor_gpu.c: new test (manufactured solution, 33x33)
• ROADMAP.md, .claude/CLAUDE.md, docs/reference/solvers.md: matrix updated

Test result: 440 GPU iters vs 479 CPU iters, max field diff 5.567e-08, L2_gpu = L2_cpu = 2.974e-02.

[Copilot]

Pull request overview

Adds a CUDA GPU backend for the plain SOR Poisson solver using an in-place Block SOR tile sweep with red-black tile coloring, closing the remaining backend gap for SOR on GPU and adding a regression test to validate convergence and agreement with CPU SOR.

Changes:

• Introduces a new CUDA Block SOR tile-sweep kernel and a new sor_gpu poisson_solver_t backend implementation.
• Wires the new GPU SOR backend into the linear solver factory/dispatcher and CUDA build sources.
• Adds a GPU SOR test (with graceful skip when CUDA/device is unavailable) and updates solver/backends documentation and roadmap tables.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
lib/src/solvers/linear/gpu/poisson_gpu_primitives.cuh	Adds the Block SOR per-thread tile sweep kernel with red-black tile coloring.
lib/src/solvers/linear/gpu/poisson_solver_sor_gpu.cu	Implements the GPU plain SOR solver backend (init/destroy/solve + factory).
lib/src/solvers/linear/linear_solver_internal.h	Declares create_sor_gpu_solver() behind CFD_HAS_CUDA.
lib/src/solvers/linear/linear_solver.c	Dispatches POISSON_METHOD_SOR + POISSON_BACKEND_GPU to the new factory.
lib/CMakeLists.txt	Adds the new .cu file to CFD_CUDA_SOURCES.
CMakeLists.txt	Adds and registers the new test_poisson_sor_gpu executable/ctest.
tests/math/test_poisson_sor_gpu.c	Adds convergence + CPU consistency test for GPU plain SOR with optional-backend skip behavior.
ROADMAP.md	Updates the backend-gap matrix and linear-solvers roadmap status.
docs/reference/solvers.md	Documents the new sor_gpu backend and GPU Block SOR characteristics.
.claude/CLAUDE.md	Updates project matrices/docs (per PR description).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.