Add meta implementation for _efficientzerotensor (#88936)
`_efficientzerotensor` is used in several backwards formulas, so its
lack of meta implementation makes those functions untracable.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88936
Approved by: https://github.com/anjali411
