Better optimization of intermediate polynomials

[Schaeff]

Before this PR, we different optimizers for the cases:

• col i = constant -> inline i
• w_0 = w_1 for two witnesses, only keep w_0
• w = lin, replace witness w with intermediate, only if lin is NOT a single witness, to avoid overlapping with the previous optimizer

After this PR, we have:

• w = lin, replace witness w (without the edge case above)
• col i = e if e is a number or a polynomial reference -> inline i

This reduces complexity in the optimizer, and also produces more optimized code, as seen in the tests.

One "breaking" change is that now when many witness are constrained to be equal, the kept on is not the first, but the last, as shown below:

col witness x;
col witness y;
x = y; 
// `y` is linear, therefore we replace the identity by
col x = y;
// `y` is a number of a polynomial reference, therefore we replace usages of x by y

x does disappear.

This works for larger sets of equal-constrained witnesses, as shown in the tests.

[chriseth]

Where do you check for "no shift"?

[Schaeff]

Actually, non-shifted is not required. Changed.

[gzanitti]

less complexity and better results, nice! 👍