Separate prover modes for high and low memory

[jonathanpwang]

The VmLocalProver can have a configuration flag for memory mode such that in high memory mode where there is no risk of OOM, it should parallelize maximally every chance it gets using rayon for work stealing. This means:

• generate_air_proof_input should be parallelized across segments. There is a stupid Send+Sync stopping it because we have some RefCell inside ExecutionSegment, but we can make a wrapper type around ExecutionSegment and unsafe impl Send + Sync just for this purpose because we know it's safe.
• proving can be parallelized across segments as well, but Engine is not threadsafe because plonky3 also uses RefCell in some DFT. If we give a Fn() -> Engine to reconstruct the engine from FriParameters in each thread, then we can parallelize.

These risk OOM since it requires much more memory, but it could improve single machine proving time.