Add GPU CI with NVIDIA Docker for OpenACC testing

[krystophny]

Problem

Currently there is no CI verification for GPU/OpenACC code. GPU regressions could go undetected.

Proposal

Add a GitHub Actions workflow that uses NVIDIA Docker to:

1. Build the interpolate library with nvfortran + OpenACC
2. Build and run GPU benchmarks
3. Verify GPU execution produces correct results (max_abs_diff = 0)

Implementation

Docker Image

Use official NVIDIA HPC SDK container:

container:
  image: nvcr.io/nvidia/nvhpc:25.11-devel-cuda12.6-ubuntu24.04
  options: --gpus all

Or build custom image with:

• NVHPC SDK 25.11+
• CUDA toolkit
• CMake, Ninja
• Required dependencies (BLAS, LAPACK, etc.)

Workflow Structure

name: GPU Tests

on:
  push:
    branches: [main]
  pull_request:
    paths:
      - 'src/interpolate/**'
      - 'bench/spline_many/**'

jobs:
  gpu-benchmark:
    runs-on: [self-hosted, gpu]  # or use GPU runners
    container:
      image: nvcr.io/nvidia/nvhpc:25.11-devel-cuda12.6-ubuntu24.04
      options: --gpus all
    steps:
      - uses: actions/checkout@v4
      
      - name: Build benchmarks
        run: |
          cmake -S bench/spline_many -B build -G Ninja \
            -DCMAKE_Fortran_COMPILER=nvfortran
          cmake --build build
      
      - name: Run GPU benchmarks
        run: |
          ACC_DEVICE_TYPE=nvidia ./build/bench_spline1d_many
          ACC_DEVICE_TYPE=nvidia ./build/bench_spline2d_many
          ACC_DEVICE_TYPE=nvidia ./build/bench_spline3d_many
      
      - name: Verify correctness
        run: |
          # Check that max_abs_diff = 0 in output
          ACC_DEVICE_TYPE=nvidia ./build/bench_spline3d_many 2>&1 | \
            grep -q "max_abs_diff     0.000000000000000"

Runner Options

1. Self-hosted GPU runner: Most flexible, requires hardware
2. GitHub GPU runners: Available in some plans (expensive)
3. AWS/GCP spot instances: Cost-effective for periodic testing

Scope

• Start with benchmark correctness verification
• Optionally add performance regression detection
• Consider adding GCC16 GPU tests once Fix GCC16 OpenACC GPU offloading runtime error #209 is resolved

Related Issues

• Fix nvfortran HDF5 complex type compatibility #208 - nvfortran HDF5 compatibility (needed for full library build)
• Fix GCC16 OpenACC GPU offloading runtime error #209 - GCC16 GPU runtime issues