[core] Refactor from_pretrained

[Cyrilvallez]

What does this PR do?

Large refactor of from_pretrained:

• Much more readable and maintainable in the future (hopefully, at least IMO)
• Tensor Parallel (TP) dispatch layer by layer, providing the best tradeoff between speed (no overhead compared to full model dispatch in one go, actually even faster) and memory (almost no memory overhead beyond the expected size of the model on the current device)
• faster model downloads (concurrent file download)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.