When hb_set_t::page_for_insert needs to insert at the end of the page_map
it ends up evaluating '&page_map[i + 1]' which has hb_vector return an
lvalue of a Crap so that nothing can be moved to its address. This turns
into issues with ThreadSanitizer on Crap when two threads modify different
hb_set_t instances. This can be avoided by using '&page_map[i] + 1'
instead.