Near-optimal tensor methods for minimizing gradient norm