[Security] Mitigate attacker lateral movement via snapshot suspend/resume system

Attacker lateral movement between nodes via snapshot suspend/resume: Model out if node isolation is enough, or if we need a different security model to restrict lateral movement of an attack in the cluster. Remember, if you get root on a node, you can mess with all the snapshots writeable by that node, which is automatic RCE on other nodes. We may need to divide the actors into N bins, and ensure we never co-schedule workloads from different bins.

Some ideas:
- Require atelet to request snapshot access tokens from the session identity broker. The session identity broker can confirm that the atelet is actively managing that session and return an access token with a CAB applied enabling access just to that snapshot.
- Alternatively, rely on per-actor identity to read and write snapshots, but only do so if actor identity is not exposed in sandboxes (it could be a local privilege escalation if a semi-compromised actor (just unprivileged UID compromise) can write snapshot memory, and then restore elsewhere with escalated privs.
- Implement ability to pin actors to specific node pools/node labels (node selector) or maybe specific warm pools that don't cross nodes. We'll have to think about whether state and user association can cross bins, beyond just a single actor/session not crossing, and how it should affect coordination between actors, e.g. if my agent needs to work with your agent but they're in different bins.