Fix Actor Memory Overhead Model #3091

We don't do a very good job accounting for actors' memory overhead today, which is a bit of a problem with the multi-actor worker model.

Today the micro-VM worker takes a flat 128MiB out of each actor's memory limit for the VMM, which is too much for small guests and too little for large ones, and nothing else accounts for what an actor costs outside its sandbox when placing it.

I propose the following API:

• Wrap SandboxRuntime in WorkerRuntime, which also has RuntimeOverhead
• RuntimeOverhead contains a fixed memory + proportional memory overhead (proportional to the actor's memory limit)

The overhead hangs off the runtime rather than the worker, because it depends on the runtime: A worker can restore actors under older runtime versions, and we may have workers that run more than one sandbox class. Snapshot and WorkerAssignment keep the plain SandboxRuntime, so the overhead isn't part of compatibility matching.

The behavior change is then:

• micro-VM stops subtracting 128MiB from the actor template's memory limit, the guest gets all of it
• scheduling starts considering the worker-published overhead model when booking an actor
• workers report their memory limit minus a small fixed amount for themselves (the worker process and in-flight snapshot writes) as capacity

Diff:

This issue is related to the next one I'll file, around mitigating OOMKill ...