refactor(compute): unify gateway restart reconciliation

[drew]

Summary

Unify graceful gateway shutdown and startup for local compute through the public ComputeDriver lifecycle RPCs. Shutdown calls StopSandbox for running-intent Docker, Podman, and VM sandboxes without persisting an explicit user stop; startup calls the idempotent StartSandbox RPC for that retained intent. Kubernetes remains cluster-owned.

Related Issue

Part of #2417

Changes

• stop running-intent Docker, Podman, and VM compute through ComputeDriver::StopSandbox during graceful gateway shutdown
• run independent shutdown stop RPCs with bounded concurrency of 16 while retaining per-sandbox lifecycle serialization
• preserve persisted lifecycle intent so gateway shutdown remains distinct from an explicit openshell sandbox stop
• terminate a gateway-managed VM driver process only after the sandbox stop sweep, and continue cleanup after individual stop failures
• reconcile retained running intent through ComputeDriver::StartSandbox before startup watch processing
• exclude explicitly stopped, stopping, deleting, and error states and leave Kubernetes compute running
• add unit coverage for phase filtering, phase preservation, driver selection, bounded concurrency, failure continuation, missing resources, and start failures
• update Docker, Podman, and VM restart E2Es to assert compute is stopped while the gateway is down and restarted afterward
• document the shared stop-on-shutdown/start-on-startup contract

Testing

• mise run pre-commit
• cargo test -p openshell-server shutdown_stop_sweep --lib
• cargo test -p openshell-server start_persisted_sandboxes --lib
• Docker, Podman, and VM restart E2E targets compile with their respective features
• OPENSHELL_E2E_DOCKER_TEST=gateway_start mise run e2e:docker
• Podman runtime E2E not run because this host has no Podman CLI
• VM runtime E2E remains blocked by the pre-existing supervisor/SSH relay readiness failure before the restart test
• mise run test: all completed suites passed except the unrelated gateway_completer_returns_empty_when_no_config, which reproduces in isolation because this host has registered system gateway configuration

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Architecture and driver documentation updated

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-2743.docs.buildwithfern.com/openshell

[github-actions]

Label test:e2e applied for f8ce804. Open the existing run and click Re-run all jobs to execute with the label set. The run will execute the standard E2E suite after building the required gateway and supervisor images once. The matching required CI gate status on this PR will flip green automatically once the run finishes.

[drew]

/ok to test 9cad488

[elezar]

Requesting changes for two lifecycle correctness issues: shutdown currently stops sandboxes serially despite the stated bounded-concurrency design, and lifecycle sweeps can issue driver RPCs from stale persisted intent. The remaining review notes are non-blocking follow-ups.

[drew]

/ok to test

Changes implemented. Re-reviewing.

[elezar]

@drew A quick summary of the changes I made while reviewing and updating this PR branch:

• Reworded the first commit footer from Closes refactor(compute): define driver shutdown and restart lifecycle contract #2417 to Refs refactor(compute): define driver shutdown and restart lifecycle contract #2417 so merging this PR does not prematurely close the broader tracking issue.
• Rebasing changed the commit IDs, but I did not otherwise alter the implementation in this PR.
• The remaining lifecycle-sweep concerns are now tracked separately: bug(gateway): drain lifecycle requests before shutdown cleanup #2826 covers draining lifecycle mutations around shutdown, and refactor: centralize persisted sandbox lifecycle traversal #2827 covers centralizing the common persisted-sandbox traversal used by the start and stop sweeps.

Those follow-ups are intentionally documented as nonblocking for this PR. The current head is 4fa5426.