We were getting some gc issues with these, I think because there
could be GC pointers behind the GC-opaque shared_ptr pointers.

Just make impl be a conservatively-allocated object so that the
collector will find it while scanning.  We still need to explicitly
call BoxIterator::gcAlloc if we store a reference (ex for BoxedEnumerate),
but this should mean that we keep these alive when they are just on
the stack (ex when we do a for-each loop over pyElements).

I don't really like how this turned out, but I think the better option
is to get rid of the indirection.