perf: investigate benefits of higher-degree constraints

[Nashtare]

We currently rely for most scenarios on log_blowup: 3, because of the overhead of the number of queries induced by smaller blowup factors.

However all constraints are currently of degree 3 or less, which feels a bit wasteful with such a blowup factor.

Immediate applications of actually allowing larger constraint degrees are:

• Poseidon2Perm table for KoalaBear (specifically): computing 2 rounds into one row, thanks to its degree-3 S-Box, would result in degree 9 constraints.
• ALU table:
• • computing a "wide" Horner accumulation, i.e. batching 4 coefficients instead of 2 in the accumulator per operation. This would be of degree 5, i.e. fitting even with log_blowup: 2
• • additional operations: we already expose a 3rd operand for the MulAdd operation. With an additional selector (preprocessed), we could support MulMul for instance
• • better encoding of selectors for fewer columns