Use a small vector optimization for the aggregate stack.

Some time ago I had tried allocating the full size up front, which
caused a performance decrease. Hopefully this hybrid is optimal.