new-regalloc performance

[Daniel Berlin]

>
> > You'll want to examine gcc.c-torture/compile/20001226-1.c.
> >
> > I'd fixed the previous big memory consumers for this test,
> > but on the branch we're back to eating 470M of memory.
>
> No surprise here, i missed a few memory leaks.
> I'm on it.

Check that, I didn't realize this test had 57169 pseudos.

It should be about half that right now.
We have an sbitmap of max_reg_num() size for each interference graph node
Among other things.

The memory usage should be cut dramatically once i remove the sbitmap per
ig_node.

--Dan