perf: gkr improvements

[gbotrel]

On my M1 laptop for bn254 gkr+mimc benchmarks (didn't test on the hpc machine, can you do it @Tabaie ? )

benchmark                 old ns/op       new ns/op       delta
BenchmarkGkrMimc19-10     14436017333     10378219291     -28.11%
BenchmarkGkrMimc17-10     3835612209      2801590584      -26.96%

benchmark                 old allocs     new allocs     delta
BenchmarkGkrMimc19-10     2822836        4238771        +50.16%
BenchmarkGkrMimc17-10     1852294        2145689        +15.84%

benchmark                 old bytes      new bytes      delta
BenchmarkGkrMimc19-10     4448614064     2635133856     -40.77%
BenchmarkGkrMimc17-10     1152920304     696349192      -39.60%

[Tabaie]

Is this a must? In the gnark API, I was creating a pool of workers for the GKR solver first and then passing the same one to the prover instead of taking it down and setting it up again, hence the pointer.

[gbotrel]

utils.NewWorkerPool returns a pointer 👍

[gbotrel]

(that being said, the pool is lightweight and it's OK to have 2 pools that are properly stopped using the "Stop" method between Solver and Prover; memory wise with stack / heap usage of the go routine of the pool that may actually work better (to be seen :))