Improved BFV Multiply (10%~)

[fionser]

• Base conversion needs many multiplications with some precomputed CRT factors.
  Thus, we can use the faster mulmod method via Shoup's trick, i.e. using the preconditioner floor(y*2^64/p).

• Same reason for multiply_poly_scalar_coeffmod.

• Performance from sealexample.cpp

clang on 2.3 GHz Intel Core i5

N = 16384 case

Average multiply: 74751 microseconds --> 65250 
Average square: 54233 microseconds --> 49981

N = 32768 case

Average multiply: 318894 microseconds --> 288964
Average square: 234881 microseconds --> 217261

g++-7 on Intel(R) Xeon(R) Platinum 8269CY CPU @ 2.50GHz

N = 16384 case

Average multiply: 89885 microseconds  —> 80355 
Average square: 66607 microseconds —> 60430

N = 32768 case

Average multiply: 380883 microseconds —> 348200 
Average square: 286745 microseconds —> 264523

[WeiDaiWD]

Thanks again for your valuable contribution. It will appear in the next release with some changes.