Add materialize_meta_tensors to reverse dispatch_model (#3866)

[discobot]

What does this PR do?

Fixes #3866.

The blocker is twofold: AlignDevicesHook.init_hook keeps the real values of offloaded weights only in hook.weights_map and leaves meta placeholders on the module, and dispatch_model wraps model.to to raise as soon as any parameter is on meta. Notably, removing the hooks is not enough: AlignDevicesHook.detach_hook only restores weights that were on a real device before dispatch, so for models whose weights went straight from the checkpoint to the offload folder (load_checkpoint_and_dispatch, from_pretrained(device_map=...)) the offloaded weights stay on meta with no public way back.

This PR adds materialize_meta_tensors(model, target_device="cpu") in big_modeling.py: it detaches all hooks (which also unwraps to/cuda), loads the weights still on meta from each offload hook's weights_map (mirroring AlignDevicesHook.pre_forward, including the int8 fp16 statistics), reties tied parameters, drops hf_device_map, and moves the model to target_device. The model is then a regular module again and .to() works for the CPU-parking round-trip described in the issue.

Four new CPU-runnable tests in tests/test_big_modeling.py cover the dispatch_model and load_checkpoint_and_dispatch flows, tied weights, and offload_buffers=True; outputs match the un-dispatched reference and .to() no longer raises or warns afterwards.

On the API shape: @SunMarc suggested a context manager or a dispatch_model flag — I went with the plain function from the issue since it is the primitive both of those can be built on (a context manager only needs to re-dispatch with the saved device map on exit). Happy to add either on top if preferred.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case. Feature Request: Add API to Materialize Meta Tensors for Device Movement #3866
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@SunMarc