Possible HashSet memory improvement

[Brian Goetz]

> Is there something I missed? Is this something that has been considered?

If memory efficiency were the only metric in the world, this would be a no-brainer.  But, by having different classes for different nodes, many many paths where the VM could prove monomorphism and thereby inline through now become polymorphic and at best could be inline cached.  

So yes -- this would be a tradeoff of memory footprint for data access costs.