[quant][graphmode] fix observer instance copy (#37185)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37185
Previously observer instances are sharing the same Tensor attributes, it is
OK if we don't do inplace operations on these attributes, but will become a problem
when people do inplace changes.
This PR uses deepcopy instead of clone_instance which will copy the tensor for each instance.
Test Plan:
.
Imported from OSS
Differential Revision: D21309084
fbshipit-source-id: afd974b0c97886fbab815e9c711c126379fe3e17