fix(sandbox): restore GPU filesystem baseline

[elezar]

Summary

Restore CUDA GPU filesystem access for Docker-backed GPU sandboxes without promoting all of /proc to full read-write policy access.

This keeps the GPU device-node baseline from #1524, but handles CUDA procfs thread-name writes with a GPU-only Landlock WriteFile exception on procfs. Non-GPU sandboxes do not receive this exception.

Related Issue

Fixes #1486

Builds on the GPU no-network enrichment fix merged in #1524.
The no-network enrichment regression is handled in #1524 and was introduced by #158. This PR addresses the follow-up GPU procfs baseline regression introduced by #910, where explicit default read-only paths prevented GPU-required baseline handling.
The GPU workload test images used for validation come from #1484.

Changes

• Keep injected NVIDIA/WSL GPU device nodes in the GPU read-write baseline.
• Stop promoting /proc into filesystem_policy.read_write; /proc can remain read-only in the policy.
• Add a Linux Landlock runtime exception that grants only AccessFs::WriteFile under /proc, and only when GPU devices are present in the sandbox.
• Cover descendant CUDA processes, such as a shell workload script that later starts deviceQuery.
• Preserve custom-policy conflicts for injected GPU device nodes that are incorrectly kept read-only.
• Update GPU sandbox policy documentation to describe the narrower procfs behavior.

Testing

• /home/elezar/.local/bin/mise run pre-commit
• /home/elezar/.cargo/bin/cargo test -p openshell-sandbox --lib baseline_tests -- --nocapture
• /home/elezar/.cargo/bin/cargo test -p openshell-sandbox --lib landlock::tests -- --nocapture
• /home/elezar/.cargo/bin/cargo clippy -p openshell-sandbox --lib --tests -- -D warnings
• Plain Docker control: docker run --rm --device nvidia.com/gpu=all localhost/openshell/gpu-workload-cuda-basic:bdaa08fb-dirty passed with OPENSHELL_GPU_WORKLOAD_SUCCESS cuda-basic
• Docker-backed OpenShell sandbox: openshell sandbox create --no-keep --from localhost/openshell/gpu-workload-cuda-basic:bdaa08fb-dirty --gpu -- /usr/local/bin/openshell-gpu-workload passed with OPENSHELL_GPU_WORKLOAD_SUCCESS cuda-basic
• A narrower /proc/self/task prototype failed the same sandbox workload with cudaGetDeviceCount returned 304, confirming the need to cover descendant CUDA processes.

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Architecture/docs updated (if applicable)

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-1522.docs.buildwithfern.com/openshell