small opt.: BN254 ML

[yelhousni]

BN optimal ate Miller loop looks like: f_{v,Q} * line1 * line2 which is implemented this way (f_{v,Q} * line1) * line2 costing 2 full-sparse multiplications MulBy034() (2 x 13M = 26M with M in Fp2). This small PR changes it to f_{u,Q} * (line1 * line2) costing 1 sparse-sparse multiplication Mul034By034() and 1 full-sparse multiplication Mul() (6M + 18M = 24M). This might be particularly interesting for multi-pairing as we would multiply the two lines 2-by-2 exploiting full sparsity.

Note: we can save 1M more when multiplying the product of lines by f as the product has always one Fp2 zero, but this is not worth it.