Making BGV/BFV noise estimates more conservative

[yspolyakov]

1. Changed the modulus switching noise expansion factor from $2 \sqrt{N}$ to $4 \sqrt{N}$.
2. Adjusted the BGV multiplication noise to use the worst-case expansion factor in the last term. In the expression for $c_{mult}$ in https://eprint.iacr.org/2021/204, the last term $r_m$ should be bounded by $\frac{N}{2t}$ instead of $\frac{\delta_R}{2 t}$ as it accounts for the overflow of messages $m$ and $m'$, which are not random.

The net effect is that the noise requirement for each multiplication level is increased by 1 bit (in both BGV and BFV).