[BUG] Normalization is missing while computing media_norm_feat during the IJB evaluation process.

First, I appreciate for your great work.

I am currently working on your IJB evalution code for face recognition and I found one suspicious thing.

https://github.com/deepinsight/insightface/blob/1bde35a620e628131141fa5c54400b7cd06f49ba/recognition/_evaluation_/ijb/ijb_11.py#L114-L147

At line 135 and 136, **I think normalization must be done right after you compute the mean feature of the multiple features having the same media_count**.

Otherwise, **for media_count == T, the more features there are, the smaller impact their mean feature will have** when you compute template_feat at line 140 since the l2-norm of the result feature vector gets smaller when we perform mean() in most cases.
As a result, it will produce uninteded template_feats.
I would be very grateful if you could explain about this.