[quant] Fix conv packed param serialization in state_dict (#52787)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52787
Current conv packed param can be serialized/deserialized with `torch.jit.save/torch.jit.load`, but
can't be saved with `torch.load(m.state_dict())/torch.save(m.state_dict())`
reason is (from James):
```
I think the issue probably has to do with the normal pickle deserialization not detecting List[Optional[Tensor]] if it doesn't witness a None in the list. IIRC this is implemented on the TorchScript side through this type tag mechanism: https://github.com/.../jit/serialization/unpickler.cpp...
```
This PR is a hack but acceptable to JIT team until a proper solution is proposed.
Test Plan:
Will be tested in next PR
Imported from OSS
Reviewed By: vkuzo
Differential Revision: D26649272
fbshipit-source-id: 4fc47a4c63e4cd1fabb404de5f0b95e127a9fca0
