feat/perf(aggregation-mode): verification & sp1 aggregation

[MarcosNicolau]

Description

While aggregating many proofs in sp1 we noticed that setting up the prover has quite the overhead (around 400ms). This pr addresses that issue by:

• Parallelising the verification in the fetching process with rayon.
• Obtaining the vk once instead of the setting up the client with the program elf every time: for this the proof pre verification was moved to the creation of the struct so that the vk is computed only once and in parallel regarded if the pre-verification is enabled or not.

Other minor things have been added:

• A pre-verification flag: to skip the verification process in the fetching process
• Logs for sp1 cuda prover have been filtered to only show up to warn levels.

Results

With a 3090 the following results were observed:

The verification in the fetching process for a batch of 40 proofs takes:

• Staging: about 1 minute
• Head: about 20 seconds

Aggregation of 150 proofs in chunks of 128 took:

• Staging: about 20min.
• Head: about 6min.

Type of change

Please delete options that are not relevant.

• New feature
• Optimization

Checklist

• “Hotfix” to testnet, everything else to staging
• Linked to Github Issue
• This change depends on code or research by an external entity
  • Acknowledgements were updated to give credit
• Unit tests added
• This change requires new documentation.
  • Documentation has been added/updated.
• This change is an Optimization
  • Benchmarks added/run
• Has a known issue
  • Link to the open issue addressing it
• If your PR changes the Operator compatibility (Ex: Upgrade prover versions)
  • This PR adds compatibility for operator for both versions and do not change batcher/docs/examples
  • This PR updates batcher and docs/examples to the newer version. This requires the operator are already updated to be compatible

[MauroToscano]

Remove the flag, and always pre verify, it has no cost and may avoid issues