[Simulation to support #702] experiment latency improvement for lookup size reduce to half

[hero78119]

Preliminary to have some data to support #702

Mock opcode number of witness & lk expression to half

riscv_add 2^20

cargo bench --bench riscv_add --package ceno_zkvm -- --baseline baseline

results

add_op_20/prove_add/prove_add_log2_20
                        time:   [1.0732 s 1.0795 s 1.0861 s]
                        change: [-33.371% -31.731% -30.166%] (p = 0.00 < 0.05)

fibonacci 2^22 max steps

cargo bench --bench fibonacci --package ceno_zkvm -- --baseline baseline

fibonacci_max_steps_4194304/prove_fibonacci/fibonacci_max_steps_4194304
                        time:   [9.9614 s 10.160 s 10.360 s]
                        change: [-23.183% -21.087% -18.757%] (p = 0.00 < 0.05)
                        Performance has improved.

other notes in this simulation

some might be inaccurate in final real case

• doesnt account for table proof m(x) SPARK + mpcs overhead, but we expect this overhead still much less than opcode prove.
• only half of witness send to mpcs.commit + mpcs.opening. But they still play in main proof witness inferring. So this part in real situation will improve more. Beside we can reduce more constrain due to limb size from 2 -> 1
• probably some opcode witness can be cut to more than half, some are less than half.