Use `torch.empty()` instead of `torch.tensor()` in `torch.nn.Parameter` (#66486)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/66486
The newly-introduced Python dispatcher mode (`__torch_dispatch__`) does not have support for `torch.tensor()` (see #64360) and this causes friction in the user experience if some `nn.Modules` use `torch.tensor()` either implicitly or explicitly.
This PR replaces calls to `torch.tensor()` in `Parameter`, `UninitializedParameter`, and `UninitializedBuffer` with an equivalent call to `torch.empty()` which serves the same purpose and is syntactically more readable.
ghstack-source-id: 140520931
Test Plan: Since no behavioral change, run the existing unit and integration tests.
Reviewed By: pbelevich
Differential Revision: D31575587
fbshipit-source-id: bd7bdeea54370f3e53dc13bd182b97d0f67146f5
