The bug was that handler::clone/handler::ha_open() call caused allocation of 
cloned_copy->ref on the handler->table->mem_root. The allocated memory could not 
be reclaimed until the table is flushed, so it was possible to exhaust memory by 
repeatedly running index_merge queries without doing table flushes.  

The fix:
- make handler::clone() allocate new_handler->ref on the passed mem_root 
- make handler::ha_open() not allocate this->ref if it has already been allocated
There is no testcase as it is not possible to check small leaks from testsuite.