Offload witness generation for LEM slots

[arthurpaulino]

The witness generation for the slots involving Poseidon hashing could be performed on GPU, which would probably provide further improvements on top of the parallelization implemented in #766.

[porcuquine]

This is stated in far too definite a form. It might or might not be helpful in the context of an optimized pipeline on a real system.

[arthurpaulino]

@porcuquine please feel free to edit 🙏🏼

[porcuquine]

I think my comment covers the nuance that was missing. In terms of substance:

• witness-generation is mostly just hashing, which we can already do in batched form on GPU (via neptune)
• but we would need to carefully modify the code to capture the intermediate values in the right order
• and this will increase GPU memory requirements significantly, as well as CPU-GPU memory bandwidth requirements

Moreover, this would put witness generation in competition with MSM for GPU usage. Currently, our best-performing proof generation shows negligible CPU usage, so the most promising approach to optimal throughput is likely better CPU utilization to minimize witness-generation time.