feat(gpu): honor device IDs in Docker and Podman

[elezar]

Summary

Honor the existing GPU device ID field for Docker and Podman GPU sandboxes without changing the protobuf API shape. Add Docker and Podman GPU e2e coverage that only runs per-device cases for CDI device IDs discovered by the container runtime, so WSL2 hosts that expose only nvidia.com/gpu=all skip index-specific cases.

Related Issue

None.

Changes

• Added a shared helper that maps the existing GPU request fields to CDI device IDs for Docker and Podman.
• Updated Docker device requests to pass explicit GPU device IDs through and keep the default all-GPU CDI request.
• Updated Podman container devices with the same explicit GPU device ID handling.
• Clarified CLI help so --gpu-device is documented as driver-specific: Docker and Podman use CDI IDs, while the VM driver uses PCI BDF or index.
• Added Rust e2e coverage for Docker and Podman GPU device selection, including default GPU requests, discovered per-device CDI IDs, nvidia.com/gpu=all, and invalid device IDs.
• Made GPU device-selection e2e compare OpenShell output against a plain Docker or Podman control container for the same CDI device request.
• Read GPU CDI device IDs from runtime info --format json discovered device entries instead of synthesizing nvidia.com/gpu=<index> from nvidia-smi output.
• Read the GPU probe image from the gateway stage in deploy/docker/Dockerfile.images, with OPENSHELL_E2E_GPU_PROBE_IMAGE available as an override.
• Removed local host nvidia-smi assumptions from the Podman GPU wrapper and Docker GPU workflow preflight.
• Switched the GPU CI workflow from the Python/k3s-oriented GPU suite to the Docker GPU e2e task.
• Repointed e2e:gpu to Docker GPU coverage and kept the previous Python GPU task available as e2e:k3s:gpu.
• Updated Docker, Podman, and testing documentation notes.

Testing

• mise run pre-commit
• rustfmt --edition 2024 --check crates/openshell-cli/src/main.rs e2e/rust/tests/gpu_device_selection.rs
• env -u RUSTC_WRAPPER cargo test --manifest-path e2e/rust/Cargo.toml --features e2e-gpu --test gpu_device_selection parse_cdi_gpu_device_ids -- --nocapture
• mise run e2e:docker:gpu on a Docker CDI GPU host after the latest rebase
• mise run e2e:podman:gpu on a Podman CDI GPU host after the latest rebase

Checklist

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

[github-actions]

Label test:e2e-gpu applied for 293b1f0. Open the existing run and click Re-run all jobs to execute with the label set. The E2E Gate check on this PR will flip green automatically once the run finishes.

[maxamillion]

I can't speak to the docker ComputeDriver changes, but this looks good to me from the podman side. 👍

[elezar]

I can't speak to the docker ComputeDriver changes, but this looks good to me from the podman side. 👍

Thanks @maxamillion. I'm trying to spin up an instance with a resonably recent podman version to test locally. One question I also had was whether there was something similar to the display of "discovered" CDI devices when running docker info in podman? (This may be useful to have if not).

[ericcurtin]

docker side looks correct FWIW