Add justfile + GPU benchmark, validate CUDA/OpenCL, fix CI TBB linking

[leonardoaraujosantos]

Summary

Adds a just-based developer workflow (modeled on the SciPP project), a GPU throughput benchmark, and fixes the Linux CI build that has been failing on every run since the GPU backends landed.

Along the way, the CUDA and OpenCL D3Q19 wind-tunnel solvers — previously authored without an NVIDIA machine and never compiled — were built and validated on real hardware (RTX 5060, Blackwell).

CI fix (the failing build)

The Linux job configured with -DCMAKE_EXE_LINKER_FLAGS="-ltbb". CMake places linker flags before the object files on the link line, so under the default --as-needed linker no symbols are yet pending when libtbb is seen and it gets dropped — every build failed with undefined reference to tbb::detail::r1::*. Adding -Wl,--no-as-needed pins libtbb regardless of order. This mirrors the justfile's configure recipe so local and CI builds link identically.

justfile

Recipe	Purpose
just bootstrap	Build + install NumPP into .deps/
just build / just test	Configure + compile + full CTest (CPU)
just gpu-detect	Probe host for CUDA / OpenCL / Metal, recommend a recipe
just gpu	Auto-enable every backend present, build + test
just cuda / just opencl / just metal	Build + run one backend's GPU test
just bench [nx ny nz steps]	CPU vs enabled GPUs, in GLUPS
just spec / just ci / just clean / just debug	OpenSpec validate / CI / cleanup / debug

Platform quirks are encoded once: Linux TBB linking (-Wl,--no-as-needed -ltbb), and CUDA built for sm_90 (SASS + PTX) so the driver JITs onto newer GPUs like Blackwell sm_120 with nvcc < 13.

Benchmark

bench/wind_tunnel_bench.cpp (behind -DCYBERFLUIDS_BUILD_BENCH=ON) times an identical wind-tunnel run per backend and reports GLUPS plus speed-up. Backend-guarded, so CPU-only hosts still build and get a CPU number.

Measured on RTX 5060 + i9-12900K (24 threads):

Backend	Precision	GLUPS	vs CPU
CPU (par_unseq)	fp64	~0.038	1.0x
CUDA	fp32	~1.15	~30x
OpenCL	fp32	~1.15	~30x

CUDA ≈ OpenCL because the D3Q19 step is memory-bandwidth-bound (~78% of the card's DRAM peak) — both hit the same roofline.

Docs

• docs/backends.md: CUDA marked ✅ validated with RTX 5060 numbers + a benchmark-reproduction section.
• README.md: new just quick-start and GPU sections; corrected the stale "CUDA/OpenCL are still stubs" claim.

Testing

• just cuda, just opencl, just gpu, just bench, just gpu-detect, just spec all pass locally on the RTX 5060 box.
• GPU solvers agree with the fp64 CPU oracle to ~1e-5 (Linf/Uin).
• The CI linker fix is verified by this PR's own CI run going green (the failure it fixes is a linker-flag issue, not covered by a unit test).