Handle graceful termination in microvm ateom

[sfunkenhauser]

Part of #23

Trap SIGTERM in the ateom, and forward SIGTERM to all application containers.
Once all application containers exit, ateom will exit itself.

[sfunkenhauser]

Benjamin Elder (@BenTheElder) PTAL

I've patched the e2e test fixes from #904 just to get a passing run, but will submit and rebase once that lands.

[Benjamin Elder (BenTheElder)]

that's in, taking a look at the other commit

[Benjamin Elder (BenTheElder)]

🤖 should-fix 🟡 – This reads ra.guestAgent with lock released, and a concurrent teardown writes it. gracefulShutdown drops the lock before calling here so a suspend can still land, and CheckpointWorkload is deliberately not gated by rejectIfDraining. Its teardownActor does ra.guestAgent.Close() followed by ra.guestAgent = nil, so the field is written while this goroutine reads it — an unsynchronised read/write that -race will flag.

The observable damage is mild: a stale pointer means signalling a closed client and logging the error, and a nil one means fifteen seconds of dialAgentRetry against a VM that is being destroyed. Both are bounded and already tolerated.

The comment above maps.Clone has the right instinct about the map; the same reasoning applies one level down to the struct it hands out. Capturing what the drain needs — the agent and workloadIDs — into a value while the lock is still held would close it, and workloadIDs is immutable after construction so only the agent really needs copying.