Consider adding some kind of "poisoned sandbox" state to prevent sandbox misuse

[ludfjig]

Certain scenario can leave a sandbox in a leaked state. For example, calling kill() on a sandbox's interrupt handle will leak any heap allocated objects in the guest. The same goes for if the guest panics, or if a host function returns an error.

We should consider introducing something like a poisoned state, which perhaps should prevent any further use of a sandbox, and require the user to restore the sandbox back into a good state via a snapshot restore, along with maybe an unsafe way to circumvent a guard like this