[Transform] Support loading random hadamards on meta device

[kylesayrs]

Purpose

• Support applying random hadamards on meta device so that they can be loaded in HF
  • [Quantization] Allow loading of transform configs huggingface/transformers#40673
  • The above PR will be able to land once this PR has been released

Prerequisites

• fix compress on meta device issue #444
• [Transform] Better dispatch support for offloaded and multi-gpu #423

Changes

• Rename _update_tied_weights to _tie_offloaded_tensors
  • The new name and comments makes its to make its purpose more clear
  • Move to apply in order to avoid redundant work
• Make _dynamic_tied_weights_keys statically defined on the transform module
  • This does not have any implications, as transformers does not check if the keys are strictly shared or even exist
• When fetching random hadamards, load them on the cpu, even if the device is meta

Testing

• Added tests model loading tests which pass after [Quantization] Allow loading of transform configs huggingface/transformers#40673 lands
• Expand test_serialization tests to check the serialized values for correctness
• Enable test_serialization_offload tests now that the latest transformers has released

[fynnsu]

Looks good! Added a suggestion below.

[brian-dellabetta]

main question regarding id() vs. .data_ptr(), but otherwise LGTM

[brian-dellabetta]

LGTM!

[fynnsu]

Nice!