Remove Symmetric Quantizer in backend (#24964)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24964
To reduce complications in quantized kernel implementation, we decided not to
have symmetric quantizer, since it can be expressed by affine quantizer,
but we will still have symmetric quantization qscheme in frontend, and user
can still specify tensors to be symmetrically quantized, while the actual quantized
Tensor represetation will only have affine quantization.
Differential Revision: D16965114
fbshipit-source-id: 0e9a5a00131878a302e211fda65a1aa427204eea