Data normalization

Normalization supports 5 modes.

freeze_bn: fixed mean and var.
```
normalize:
  type: freeze_bn
```
solo_bn: seperately counting mean and var by the single GPU.
```
normalize:
  type: solo_bn
```

pt_sync_bn: synchronizing mean and var by multiple GPUs by pytorch.

normalize:
  type: pt_sync_bn
  kwargs:
    group_size: 8

gn: Group Normalization

normalize:
  type: gn
  kwargs:
    num_groups: 32

caffe_freeze_bn: using the frozen bn preloaded from caffe.
```
normalize:
  type: caffe_freeze_bn
```