K2-Think: Teaching a 32B Model to Reason Like the Big Ones