Multiplication with 8-decdig batch

[tompng]

Perform multiplication with 8-decdigs * 8-decdigs batch. This will reduce carry operation by 1/4, and may enable vectorization.

Benchmark

irb(main):001> measure
TIME is added.
irb(main):002> x = BigDecimal('111111111'*num_decdig); t = 100000.times{x*x};

Smaller is better

num_decdig	time(master)	time(this branch)	memo
99	4.388227s	0.971541s	4.5 time faster
50	1.133961s	0.276888s	4.1 times faster
20	0.216531s	0.076732s	2.8 times faster
10	0.080238s	0.041933s	1.9 times faster
4	0.018200s	0.017952s	same
1	0.0147669s	0.016004s	slower

Optimal threshold of using NTT multiplication and Newton-Raphson division has changed because multiplication gets about 4 times faster (when num_decdig is near threshold).