[Pre Neuron Inf Cache system]Support neff/weights decoupling

[JingyaHuang]

With inline_weights_to_neff argument, we are now able to decouple the weights and neff graph during the compilation. It means that we can set up a caching system for reusing shared compiled neff graphs and load with weights from different checkpoints during the inference. This could largely save the compilation time which could take up to hours.

• Support inline_weights_to_neff in Neuron exporter
• Replace weights function that works for Optimum wrapped models
• Support replacing weights in the modeling
• Tests

[Next step]
Initial caching for inference of encoder models

• Set up hashing for inference
• Set up caching mechanism based on current caching of models with independent neff files (optimum-cli for creating public/private cache)
• Checker & replace weights on the modeling from_pretrained when export=True and cache exist.
  • Exist (local / remote) -> load precompiled neff + replace weights with current checkpoint
  • Non-exist -> Compile from scratch and maybe cache

[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.

[dacorvo]

LGTM, but you have a typo in your tests.