Fixed-Budget Best-Arm Identification with Heterogeneous Reward Variances