pvf: better approval job scheduling

[sandreim]

We should investigate if under heavy load, the node can avoid wasting CPU time by skipping executing candidates included in finalised blocks.
In theory it should be possible that the pov recovery completes after the inclusion block was finalised, in which case we can simply skip executing the PVF.
Suggested in #5616 (comment)

[alexggh]

Thought about this idea as well, one thing to have in mind is that if B1 is finalized on node N1 doesn't not mean it is also finalized on node N2, so whatever solution we pick we need to make sure we don't cancel approvals that other nodes are still waiting for.

[burdges]

I doubt this matters. If we announce our approval assignment then finality is blocked on us. If we take so long we no-show, then in expectation 2.25 others assignee themselves, and finality is blocked on them, or us, whichever is faster. If they're faster but we're online and still checking, then we've either hit some fluke of networking (unlikely), or else our node has too little bandwidth or too little processing power (likely).

As polkadot's developers we're primarily worried about this last scenario, which means replacing the slow node in the network, not improving its internal performance metrics when its screwing up ours. I think polkadot-fellows/RFCs#119 plays an important role here, but the slow node wasting its resources maybe speeds up it losing nominators & elections.