Make the unit of preemptive termination controllable

[dead-claudia]

I'm thinking this could employ some rules to reduce the size of the unit of preemptive termination, as well as allowing users to artificially reduce it further as desired:

• If an agent lacks a reference to a shared array buffer, it terminates that agent only.
• If an agent has a reference to a shared array buffer, all agents are terminated by default.
• Agents can opt on a per-agent basis to remain alive when another agent referencing a shared array buffer it also holds a reference to terminates. In this case, the opt-in makes it clear that the agent should be prepared to deal with potentially inconsistent state. (There are mechanisms for ensuring consistency in the face of potential sudden preemption - operating systems already necessarily use these for reasons like power failure tolerance.)
• In all cases, the SAB remains alive and usable as long as agents with references to it retain their references.

This would be able to strike a compromise and avoid the issues detailed in https://github.com/tc39/proposal-oom-fails-fast#can-the-uopt-be-finer-grain-than-an-agent-cluster.