Improve QuickSilver performance

[themighty1]

Currently, our QS impl accumulates all the triples from all vm calls which leads to memory blowup.

E.g. proving 30KB of AES encrypted data requires ~12 mil triples or 12*48=~600MB of memory.

Additionally, the code initializes the memory often. The following hot path contributes to 0.8s of the runtime with 12 mil AND gates:

mpz/crates/zk-core/src/check.rs

Lines 41 to 46 in 16c2e7a

	pub(crate) fn reserve(&mut self, n: usize) -> usize {
	let idx = self.triples.len();
	self.triples.resize_with(idx + n, Default::default);
	self.adjust.resize_with(idx + n, |_| Default::default());
	idx
	}

We can avoid the memory/compute overhead by treating each individual vm call as a separate circuit and using its adjust bits for its local transcript which gets hashed to derive chi.

We would need to change the logic, so that upon executing the call, the prover/verifier immediately pre-compute the values for the consistency check and discard the triples.

[sinui0]

Related #200