accelerate wit_infer_by_expr via grid-strde-loop & vector pool

[hero78119]

related #772

This PR re-design wit_infer_by_expr to try grid-strde-loop traverse + reuse a simple vector pool to make vector reuse across closure of expression evaluation. This design is aiming to avoid massively interface change.

highlight of change

• introduce a simple vector pool to maximize vector reuse during expression evaluation, as we can't just allocated a simple vector and mutable it, due to the expression degree > 0, thus inevitable there will be intermediate poly.

  the evaluation of read/write/lookup expression also changed to sequential, so can maximize the vector pool utilization.

• refactor wit_infer_by_expr to use grid-strde-loop style and maximize cpu cache hit rate. vector are updated concurrently with SyncUnsafeCell, which require nightly feature #![feature(sync_unsafe_cell)].

  SyncUnsafeCell is just an UnsafeCell in stable, except it implements Sync if T implements Sync.

benchmark

with command

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

on ceno server (16 cores, 32 HT) fibonacci e2e latency reduce ~13%

fibonacci_max_steps_1048576/prove_fibonacci/fibonacci_max_steps_1048576
                        time:   [4.1685 s 4.1951 s 4.2220 s]
                        change: [-15.436% -13.927% -12.397%] (p = 0.00 < 0.05)
                        Performance has improved.

[hero78119]

There are many code redundant since I still do the experiment and see what the best design and avoid massive change its done and cleanup

[hero78119]

It seems I am test on the wrong branch, will update it once benchmark result is confirmed