Double-precision cubic towering + pairing

[mratsim]

This PR continues the work of double precision towering

• Double-width tower extension part 1 #72
• FpDbl revisited #144
• Rework towering #148
• Double-Precision towering #155

Hopefully we get so efficient that even BLS24 curves become interesting.

The first commit improves Fp12 multiplication by 5%, which tranlates to about 4% speedup on pairings.

Fp12 mul, before

Fp12 mul after

[mratsim]

PR is finished, we obtain a 3% boost in pairings overall.

• The Miller Loop requires a specialized double-precision sparse codepath
• The final exponentiation requires optimized Fp12.
• Some lazy additions were not used but could improve perf (by hundreds of cycles on Fp12). They seem to work for BLS curves, but not for BN curves, is it because of the M-Twist/D-Twist or the fact that there is at least 3 spare bits and so we can store 3p on Fp or 6p on FpDbl, who knows.

Before

After