heapsort shootout rewrite

[Stephen Weeks]

> I don't see more than 10% either way, but the code is more readable.  Here is
> the way I would write it (noopt.sml) and then a tweak to avoid fetching array
> elements  more than once (opt.sml).  The latter only seemed to buy 3% for the
> 80,000 element case, so it really isn't worth any thing.

Here are the times that I see for the original submission and your two optimized
versions, on both the released MLton version and our internal version.

		20000906	internal
original	11.2		5.1
noopt		 9.9		5.0
opt		 7.8		4.9

Based on this, I think I'll submit noopt.  The real moral is we need to get the
release out :-)