Supervises one model version's worker pool.

One of these per loaded {name, version} with an :exclusive backend. Shared-concurrency models start no workers, so no supervisor is started for them either.

Scoping the pool to a single model version is what keeps failures isolated: a backend that crash-loops exhausts its own restart intensity and takes down its own model instance, while every other model keeps serving. A single global pool would let one bad model's restarts either starve or kill unrelated ones.