cuda: Add compile test only for CI

[StormBytePP]

Summary

Adds a CUDA compile check to the libvmaf GitHub Actions workflow so that -Denable_cuda=true is exercised on every push/PR.

This implements the request in #1576: a lightweight CI job that installs the CUDA toolkit + nv-codec-headers, configures with Meson, and builds/installs. Runners do not have an NVIDIA GPU, so device tests are not run; the goal is to catch configure/compile regressions early.

Closes #1576

Build directory (out-of-tree, CUDA jobs only)

Only the CUDA matrix entries configure the build in a temporary directory outside the source tree (e.g. $RUNNER_TEMP/libvmaf_build). All other jobs keep the existing in-tree layout (libvmaf/build), so tests, tox, and packaging behave as on master.

Out-of-tree for CUDA is intentional:

• With the current Meson CUDA rules, several nvcc/clang custom targets use relative include paths such as -I ./src, -I ../src, etc. Those paths work when the build directory is a subdirectory of the source root, but break when the build dir is fully outside the tree.
• Using an out-of-tree build for the CUDA jobs makes that class of failure visible instead of hiding it behind an in-tree layout.
• CUDA: Fix compile, SIGSEGV and tests #1573 fixes those relative includes (among other CUDA compile/test issues) by using absolute paths based on meson.current_source_dir() / the build directory, and adjusts related test wiring. Once CUDA: Fix compile, SIGSEGV and tests #1573 is merged, the CUDA CI jobs should stay green for out-of-tree builds as well.

So this workflow both guards CUDA support and documents why out-of-tree builds matter for the CUDA path, without disrupting the rest of the matrix.

What the CUDA jobs do

Two matrix entries on ubuntu-latest (x86_64):

Job	Host compiler	Device compiler
linux-x86_64-gcc-cuda	GCC	nvcc (-Denable_nvcc=true)
linux-x86_64-clang-cuda	Clang	clang (-Denable_nvcc=false)

Both:

• Enable CUDA with -Denable_cuda=true
• Install a compile-only CUDA toolkit (no driver/runtime metapackages that pull in nvidia-open and fail on CPU-only runners)
• Install nv-codec-headers (FFmpeg) so ffnvcodec/dynlink_cuda.h is available
• Run meson setup + ninja install only (no GPU tests)
• Use an out-of-tree build directory ($RUNNER_TEMP/libvmaf_build)

Existing non-CUDA jobs are unchanged (in-tree build, tests, tox, packaging).

Notes

• Covering both nvcc and clang as device compilers matches the two paths exposed by -Denable_nvcc. Toolkit and Clang versions on the runner may need occasional bumps when Ubuntu/NVIDIA package sets change.
• MinGW/MSYS2 CUDA was not added here: cross-compiling CUDA for Windows from Linux is awkward and not a good fit for this check. Native Windows/MSVC CUDA coverage can follow once MSVC support is merged.
• CUDA toolkit installation: The workflow currently uses the third-party action Jimver/cuda-toolkit purely for convenience (precise version control + compile-only packages).
  If the maintainers prefer to avoid external actions, this can be trivially replaced with a manual installation via apt (or the official NVIDIA runfile) on Ubuntu. Just let me know and I’ll update it.

Test plan

• Non-CUDA matrix jobs still configure, build, test, and package as before (in-tree libvmaf/build)
• GCC + nvcc CUDA job installs toolkit + headers and completes meson setup + ninja install (out-of-tree)
• Clang + clang-device CUDA job installs toolkit + headers and completes meson setup + ninja install (out-of-tree)
• After CUDA: Fix compile, SIGSEGV and tests #1573: confirm out-of-tree CUDA builds remain green