Take two of the math128 bits.. 

Changes since last time:
 - Added gcc __int128 versions for where it works
 - Added Andy's mul_u64_u32_shr() due to popular demand :-)
 - Added {shl,shr}_u128 as requested by hpa (although I suspect he'll
   use Andy's fancy function).
 - Added a U128_INIT() constant initializer.
 - Renamed mult_u128 to mul_u64_u64, hpa complaned the name was
   misleading
 - Fixed some of the code generation issues with add_u128 and
   mul_u64_64.

I've also assumed we're all actually ok with adding this functionality,
so unless someone hollers I'll push this for inclusion.