Also, use a cloner instead of unmarshalling when wrapping an object.

Increasing `numWorkers` doesn't seem to help at all (not supprising). Really, we
could probably drop to NumCPUs (instad of 2x that) and we'd still be fine.

More than that doesn't help. +1 helps for the case where a goroutine holding a
worker gets unscheduled.

IIRC, pools are cleaned every GC cycle. However, by benchmarks aren't
actually *faster* with this change so I figured we might as well go for it. Our
issue here is really throughput, not latency, so even *if* these pools are
cleaned on GC, it's still useful.

We don't need to mutate, just replace.

Make sure we *actually* serialize/deserialize. Addresses @warpfork's CR.