quant norm layers: move scale + zp to buffers (#52861)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52861
Currently, scale and zp in these layers are not buffers, which
means they do not get saved to the state dict. Movin them
into buffers to allow people to properly use state_dict.
Note: this is a redo of https://github.com/pytorch/pytorch/pull/45313,
with BN taken out. Not doing this for BN because it has dependencies on existing
behavior. We should clean it up eventually.
Note: not handling BC because it's 100% broken now, so there is
no practical value in handling BC.
Test Plan:
```
python test/test_quantization.py TestPostTrainingStatic.test_normalization
```
Imported from OSS
Reviewed By: jerryzh168
Differential Revision: D26671761
fbshipit-source-id: 7615b1dd0d1ae88eeff8b1d150f3846815dc2bc9
