Investigate decimal128_fast multiplication slowdown

[mborland]

Takes about twice as long for decimal128_fast multiplication vs that of decimal128. Maybe because we are only using emulated types in the former?

[mborland]

Can the significand of the type be unsigned __int128 when available and only our emulated u128 when necessary? This should provide an immediate speedup.

[mborland]

In decimal128 we can check if the result of the two significands will be less than or equal to 128 bits and then use simplified multiplication if that's the case. For decimal128_fast the significands are always normalized so we are forced to use 256-bit multiplication which will always be slower. Removing all the trailing zeros of both significands to perform a similar comparison is quite expensive.