Fix potential code caching/re-ordering issue in __search_insert_append

if ((ins = WT_SKIP_LAST(ins_head)) == NULL)

        return (0);

key.data = WT_INSERT_KEY(ins);

key.size = WT_INSERT_KEY_SIZE(ins);

WT_RET(__wt_compare(session, collator, srch_key, &key, &cmp));

if (cmp >= 0) {

        /*

         * !!!

         * We may race with another appending thread.

         *

         * To catch that case, rely on the atomic pointer read above

         * and set the next stack to NULL here.  If we have raced with

         * another thread, one of the next pointers will not be NULL by

         * the time they are checked against the next stack inside the

         * serialized insert function.

         */

        for (i = WT_SKIP_MAXDEPTH - 1; i >= 0; i--) {

                cbt->ins_stack[i] = (i == 0) ? &ins->next[0] :

                    (ins_head->tail[i] != NULL) ?

                    &ins_head->tail[i]->next[i] : &ins_head->head[i];

                cbt->next_stack[i] = NULL;

        }

        cbt->compare = -cmp;

        cbt->ins = ins;

        cbt->ins_head = ins_head;

        *donep = 1;

}