Build CPU + Thrust + OMP-offload from a single install when NVHPC is present

[hfwen0502]

Summary

Previously SBD_BUILD_BACKEND=auto built only CPU + Thrust, and the OpenMP target-offload backend (_core_gpu_omp_offload) had to be installed by itself in a separate venv, on the assumption it would conflict on the OpenMP runtime with the other backends.

That assumption no longer holds. On a GPU box all three extensions compile with NVHPC nvc++ and therefore link the same OpenMP runtime (libnvomp) — CPU via -fopenmp (nvc++ treats it as -mp), Thrust via -mp -cuda, offload via -mp=gpu. They co-load in one process without conflict, so a single pip install can now provide all three.

Changes

setup.py

• auto: build CPU + Thrust + OMP-offload when NVHPC is present; CPU only (g++/clang) when it is not.
• Add SBD_BUILD_BACKEND=all (builds all three; errors if nvc++ is missing).
• When the CPU backend is built alongside any GPU backend, route the whole build through nvc++ up front so the CPU extension also links libnvomp.
• macOS (clang) and Linux-without-NVHPC (g++) paths are unchanged. Backend selection is keyed solely on nvc++ availability — no GPU-hardware probe — so a GPU-less build node with the SDK still builds the GPU backends (build-here / run-on-compute-node stays supported).

README.md

• Rewrite the install section for the single-venv unified build (one venv, one pip install).
• Update the SBD_BUILD_BACKEND table (auto/all) and correct the backend co-load description and verify output.

Testing

Submodule pinned at the current vendor/sbd-upstream (21f73b4); no submodule change in this PR.

	Linux + 8×H100 (NVHPC 26.3)	macOS (clang)
auto builds	_core_cpu + _core_gpu_thrust + _core_gpu_omp_offload	_core_cpu
available_backends()	['cpu', 'gpu', 'gpu-omp'] (one process)	['cpu']
h2o (1e-3) energy	-76.2359465468 — all three backends agree bit-for-bit	-76.2359466293 (within 1e-4 Davidson tol)

On the GPU box the full SQD loop also runs end-to-end and converges.

[hfwen0502]

the solution provided is wrong. Not valid. still need to separate out the runtime between omp offload and cpu/thrust.