Design a way to "pin" the result of floor planning

[str4d]

Circuits need to specify the floor planner that they use, because the floor planner has explicit power to alter the structure of the circuit (and thus the verifying key). Once a circuit has been completed and deployed, re-running the floor planner is pointless busywork, as the verifying key is pinned (as input to the Fiat-Shamir transform). But currently every user of a deployed circuit runs their chosen floor planner (e.g. floor_planner::V1 for Orchard) every time a proof is created.

We should have a way to encode the layed-out circuit in some way, and then feed that into the prover so that Circuit::synthesize is called precisely once (to assign the per-instance witnesses), instead of e.g. twice for floor_planner::V1. This could probably just become part of the ProvingKey, both in-memory (i.e. each user of a circuit computes the layout once on startup and then caches it for all proofs), and eventually stored in a serialized form (#443).

[str4d]

This pinned format might also end up being how other implementations of Halo 2 interoperate (e.g. some other frontend that can produce Halo 2 circuits could pass that data to halo2_proofs as a backend), but that's a secondary concern (that may end up being solved in other ways).

[daira]

The description of the duplicate #117 was:

I think it should be possible to dump an optimized configuration + floor plan in serialized form (e.g. as Rust code that constructs it), and load it back in rather than doing the region sizing / floorplanning / optimization at run time. For Pollard we can use a single combined core and do such optimizations by hand if they're applicable. But I'd like to make sure that the API is such that they could be done automatically.