SML numerical benchmark and MLton

[Stephen Weeks]

> Amusing.  Other than the unboxing, why is MLton so much faster?

No idea.  There's a little bit of higher order stuff (fold, app,
etc.), but I get the feeling it's mostly unboxing doubles and
monomorphisation.  Could also be some uncurrying and some gcc stuff
(register allocation etc.) because it looks like pretty vanilla code.

mlprof indicates that all the time is spent in main, so MLton was able 
to turn everything into loops.

Still, I don't have any explanation why this benchmark is so much
better than say, ratio-regions or tsp.