Reliably catch compile/prove discrepancies

[mitschabaude]

Follow-up to #624

• Investigate why a discrepancy between compilation & proving caused an out of bounds error during the compute_witness in the Pickles wrap prover
• Exactly identify what in that bug was the underlying problem, and what the course of downstream effects, at a low level
• Find a way to identify this error earlier in the process, such that the error is more closely related to the cause
• Reliably communicate the underlying problem to the user

[mimoo]

Interesting, reading #498 it seems like we're trying to access a witness value as part of exists (witness generation) or as_prover (debug). I think without better errors this will be hard to debug, so if this not urgent perhaps we can wait on getting better errors first

[mitschabaude]

This again cropped up as an important issue as several zkApp testers ran into these kinds of issues and they proved to be extremely hard to debug. A simple first step would be to reliably tell the user if there is a compile / prover discrepancy. It seems to do that we'd need to create a constraint system in the prover and return its digest, in a way that it can be used if the proof fails to identify this as a problem