Skip superfluous storage allocations while constructing meta tensors (#65331)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65331
ghstack-source-id: 148862595
This is a performance optimization for the use case:
```
tensor = torch.tensor(<large_data>, device='meta')
```
where the current implementation requires a superfluous memory allocation on CPU even though the target device is a meta.
Test Plan: Run existing tests since no behavioral change is introduced.
Reviewed By: ezyang
Differential Revision: D31055036
fbshipit-source-id: 04d6c13594a71fc65bf2fbd567ee71833a879851
(cherry picked from commit 489d0a151a5fc4f5a0d8e3e65897bf7d02affe4b)
