[PyTorch] One less refcount bump in linear() (#54936)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54936
Another case where we can use `MaybeOwned<Tensor>` to save a bump at a small cost.
ghstack-source-id: 125218488
Test Plan: Existing CI
Reviewed By: ngimel
Differential Revision: D27421117
fbshipit-source-id: 16bb31ec38817be1f889360e2abfd0d9596e2943
