Limit number of tasks spawned for client proof-of-work

Currently we spawn a unlimited number of tasks for doing client proof-of-work computations. We should put them on a more limited threadpool. We might do this with spawn_blocking (tokio, async-std), or we might want to put it on a threadpool restricted for use only for PoW calculations.

As per @nickm's comment on !2026 (closed), this is a blocker for making client PoW non-experimental.