[Feature]: AutoDeploy: new load hook system

🚀 The feature, motivation and pitch

Currently, AutoDeploy load hooks are part of PyTorch's built-in load hook system inside nn.Module.

Let's write a custom load hook system that we have more control over for the following reason:

1. PyTorch's built-in load hook system is hierarchical (i.e. the order is determined by both the order of registration and the submodule hierarchy, see ticket #10348) This makes ensuring a constant order confusing. We should instead focus on having a flat load hook system where load hooks are always executed on the state dict / root module and order of execution is purely determined by the order of registration
2. Execute load hooks on CPU during pre-loading. This ensures we only allocate GPU memory we actually need for the weight and ensures better control over the memory management

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.