make sym sizes be computed lazily (#82233)
### Description
Creating size nodes proactively for each tensor is leading to increased memory pressure as hold strong pointers to tensor data.
### Issue
[<!-- Link to Issue ticket or RFP -->](https://github.com/pytorch/pytorch/issues/80942)
Creating
### Testing
<!-- How did you test your change? -->
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82233
Approved by: https://github.com/wconstab
