fix(compute): allow stopping errored sandboxes for recovery

[shiju-nv]

Summary

An infrastructure startup failure can leave a retained sandbox in Error, with both stop and start refusing recovery even after the underlying problem is fixed. Allow explicit stop from that state so operators can use the existing start path without deleting the sandbox or its persistent workspace. Provisioning timeouts keep their existing cleanup-then-start recovery path.

Related Issue

Closes #3308

Changes

• Accept ordinary Error states in stop, while preserving pending-operation and timeout cleanup checks.
• Tell direct-start callers to stop the sandbox before retrying.
• Cover startup-failure recovery, renewed authentication and configuration admission, driver error reasons, pending operations, timeout cleanup, and failed-stop retries.
• Document the recovery commands and the provisioning-timeout exception.

Testing

• Checks appropriate to the affected code and behavior pass
• Unit tests added/updated
• E2E tests added/updated (local Docker reproducer executed; no new repository E2E test)

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Architecture docs updated (sandbox lifecycle documentation)

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-4231.docs.buildwithfern.com/openshell