big: improved computation of "karatsuba length" for faster multiplies

big: improved computation of "karatsuba length" for faster multiplies

This results in an improvement of > 35% for the existing Mul benchmark
using the same karatsuba threshold, and an improvement of > 50% with
a slightly higher threshold (32 instead of 30):

big.BenchmarkMul           500	   6731846 ns/op (old alg.)
big.BenchmarkMul	   500	   4351122 ns/op (new alg.)
big.BenchmarkMul           500	   3133782 ns/op (new alg., new theshold)

Also:
- tweaked calibrate.go, use same benchmark as for Mul benchmark

R=rsc
CC=golang-dev
https://golang.org/cl/1037041