TF safetensors reduced mem usage (#24404)
* Slight comment cleanup
* Reduce peak mem usage when loading TF-format safetensor weights
* Tweak the PyTorch loading code to support lazy loading from safetensors
* Pass safe_open objects to the PyTorch loading function
* Do GPU transposes for speed
* One more tweak to reduce peak usage further
* One-line hasattr
* Fix bug when there's a shape mismatch
* Rename state_dict in the loading code to be clearer
* Use TF format everywhere for consistency
