Summary

Replace GPU-specific sandbox request fields with PlacementRequirements / GPUSpec, and add optional GPU count placement requests while preserving the default GPU request shape.

Related Issue

None

Changes

• Added PlacementRequirements with an embedded GPUSpec to public and driver protos.
• Added optional GPUSpec.count, mutually exclusive with explicit device_id values.
• Added openshell sandbox create --gpu-count, with CLI validation for count > 0 and --gpu-device conflicts.
• Validated GPU count/device-ID mutual exclusion at the gateway and preserved count in driver request translation.
• Honored GPU count in Kubernetes by setting nvidia.com/gpu to the requested count.
• Kept VM support limited to one GPU, and rejected count requests for Docker and Podman until those drivers support count semantics.
• Preserved image-name GPU auto-detection and --gpu as a present GPU request with no device IDs and no count.
• Updated unit tests and user/architecture docs for the new placement behavior.

Testing

• mise run pre-commit passes
• cargo check -p openshell-core -p openshell-server -p openshell-driver-docker -p openshell-driver-kubernetes -p openshell-driver-podman -p openshell-driver-vm -p openshell-cli
• cargo test -p openshell-cli gpu
• cargo test -p openshell-server -p openshell-driver-docker -p openshell-driver-kubernetes -p openshell-driver-podman -p openshell-driver-vm -p openshell-core gpu --lib
• Unit tests added/updated
• E2E tests added/updated (not applicable; e2e coverage was already merged separately)

Checklist

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