feat(sandbox): add explicit suspend and resume lifecycle #4763

User Story

When validating OpenShell sandbox lifecycle, I need to suspend and resume a supported workload without restarting its execution, while preserving gateway authentication and lifecycle boundaries.

Problem Statement

OpenShell needs an explicit gateway lifecycle for retaining a sandbox workload across suspension, restoring it under a new generation, and reporting unsupported driver semantics consistently.

Impact / Why This Matters

Stopping and starting a sandbox discards live workload state. An explicit suspend/resume workflow lets operators pause a supported workload and continue it later.

Proposed Design

Expose sandbox suspend and resume through the gateway, CLI, and SDKs. Drivers advertise support. Suspension makes a sandbox unavailable while its backend retains the workload; resume creates a new runtime generation and restores it. A strict restore failure preserves the ability to retry. Explicit start, stop, and delete discard retained state.

The concrete V8 implementation and manual counter demonstration are tracked separately in #4857, on top of the lifecycle contracts in #4765.

Suggested UX

openshell sandbox suspend my-sandbox
openshell sandbox resume my-sandbox
openshell sandbox stop my-sandbox
openshell sandbox start my-sandbox

Acceptance Criteria

• Supported sandboxes transition through Suspending and Suspended and resume into a fresh runtime generation.
• The driver/backend contract supports retained workload progress without replaying initialization.
• Failed restore launches preserve the ability to retry when retained state remains available.
• Explicit start, stop, and delete discard retained state.
• Unsupported compute drivers reject suspend/resume semantics.
• CLI and SDK clients expose the new lifecycle operations and phases.
• Lifecycle tests cover successful transitions, failed resume recovery, and exits racing suspension.

Alternatives Considered

Stop/start remains the clean-boot workflow. Backend-specific retained-state implementations are exercised separately from the shared lifecycle contracts.

Context

Related: #4352 discusses suspension initiated by a backend and supervisor reconnect behavior. This issue covers explicit gateway-driven suspend/resume. The dependent V8 example is tracked by #4857.