* Change CycleFold approach:

Instead of having a single CycleFold circuit that checks the 2 forign
scalarmul of the main circuit instances, now there are 2 separated
CycleFold circuits each of them checking a single foreign scalarmul.
Increasing the number of constraints of the AugmentedFCircuit, but
reducing the number of constraints in the CycleFold circuit, which will
translate into reducing the number of constraints in the Decider
circuit.

* CycleFold circuits checks in AugmentedFCircuit:

- update NonNativeAffineVar to work with NonNativeFieldVar directly
  instead of FpVar comming from NonNativeFieldVar.to_constraint_field()
- include in AugmentedFCircuit intermediate steps inbetween CycleFold
  circuits, and update the internal checks of the CycleFold circuits

Pending to document the new CycleFold circuits approach and better
variable namings, rm unwraps, etc

* matrix_vec_mul_sparse gadget: skip value * v[col_i] mul when value==1

Saves a notable amount of constraints since there is a notable amount of
1 values in R1CS matrices.

* Reuse computed vector of U_i

Reuse computed vector of U_i, saving 4k constraints in AugmentedFCircuit.

* fixes post last rebase to main

* rm test_augmentedfcircuit since it is already tested in test_ivc (and is a slow computation)

* rm dbg!()

* small fixes after last main rebase