Remove feature gate and improve benchmarks

[ggwpez]

To measure the impact of optimizing the proof batching more precisely.

Changes:

• Remove small-ring feature gate but instead use Rust generics
• Run benchmarks repetitively to get more accurate results
• Add benchmark script for Parity ref hardware for comparable numbers

Finding: Commitment encoding for the full ring is very slow; 2572 ms on reference hardware 😢

Small ring on reference hardware:

* PCS params decode (1000 reps): 13 ms
* Open (10 reps): 20 ms
  Commitment size: 386867 bytes
* Commitment encode (100 reps): 0 ms
* Commitment decode (5 reps): 96 ms
* Create: 68 ms
  Proof size: 788 bytes
* Start members (1000 reps): 0 ms
* Push members (1000 reps): 13 ms
* Finish members (100 reps): 0 ms
* Validate (100 reps): 12 ms
* Alias (1000 reps): 0 ms

Big ring:

* PCS params decode (1000 reps): 40 ms
* Open (10 reps): 195 ms
  Commitment size: 12322611 bytes
* Commitment encode (100 reps): 13 ms
* Commitment decode (5 reps): 2572 ms
* Create: 790 ms
  Proof size: 788 bytes
* Start members (1000 reps): 0 ms
* Push members (1000 reps): 12 ms
* Finish members (100 reps): 0 ms
* Validate (100 reps): 29 ms
* Alias (1000 reps): 0 ms

[ggwpez]

Why? FRAME is single threaded so a spin lock wont slow us down.